APLACEGRAPHDATABASEASAQUALITATIVEHUMAN PLACEKNOWLEDGEBASE
hao chen
A thesis submitted in total fulfillment of the requirements for the degree of Doctor of Philosophy
Jan 2019
Department of Infrastructure Engineering The University of Melbourne
[ April 29, 2019 at 20:02 – classicthesis version 5 ] Hao Chen: A Place Graph Database as a Qualitative Human Place Knowl- edge Base, © Jan 2019
[ April 29, 2019 at 20:02 – classicthesis version 5 ] ABSTRACT
Place, as a human spatial concept, has been studied extensively in the fields of philosophy and psychology. Recently, place-based research has received an increasing amount of attention in the domain of GI- Science, and the motivation is to study how people cognitively rep- resent and communicate about place knowledge, in order to design place-based information systems to facilitate human-computer inter- actions. The motivation is strengthened as the need for place-based information systems is growing, facing the emergence of ubiquitous computing with users no longer necessarily be trained GIS experts. This PhD research is part of a project that aims at designing a place-based geographic information system in order to make human place knowledge digestible by computers. This thesis takes natural language place descriptions as the source input of study, and focuses on four research aspects as four major tasks: place description knowl- edge modelling, qualitative spatial reasoning, place georeferencing, and knowledge querying. Place descriptions occur in everyday ver- bal communication as a way of encoding and transmitting knowledge about places between individuals, and thus they reflect the way that people mentally represent and communicate about place knowledge. Place descriptions typically provide qualitative reference systems for describing the locations of places using place references and spatial relations, as well as provide other non-spatial information such as place semantics and place characteristics. The abundance of natural language place descriptions suggests complementary approaches for human place knowledge representation. The first task looks at how the human knowledge about places ex- tracted from place descriptions can be modelled in an information system. The second task is based on that people typically use place references and qualitative spatial relationships to describe the rela- tive locations of places. Such qualitative relational knowledge could be used for inferring new relational knowledge, as well as for reason- ing to maintain the logical consistency among the stored knowledge in the information system. The third task focuses on developing a methodology for georeferencing all places captured in the system in order to locate them on the map. Last, from a knowledge base per- spective, the developed system should also support querying of the stored collective knowledge in order to utilize such knowledge in ap- plication scenarios. The scientific contributions of this thesis include novel developed approaches addressing the four tasks, specifically: a) a designed place graph database model for the efficient storage and query of place knowledge extracted from place descriptions, and its superiority is demonstrated by comparing to a baseline model from the literature; b) a spatial reasoning framework for maintaining local relational con- sistency in transactions happened within a place graph database; c) a
iii
[ April 29, 2019 at 20:02 – classicthesis version 5 ] multi-step georeferencing approach for locating all places in a place graph database, with sub-contributions including a novel clustering algorithm, spatial relationship search space models, a weighted multi- value gazetteer matching approach, and an approach to derive den- sity surfaces to represent the approximate locations of places; d) a place knowledge query framework for different types of structured queries based on graph traversal, including a dialog-based approach developed for query contextualization. A place graph database man- agement system has been developed that integrates the above ap- proaches, together with other functionalities such as graph visualiza- tion and place mapping. Such a place graph database can be regarded as a qualitative human place knowledge base constructed from collec- tive place descriptions. keywords: place description, place graph database, place knowl- edge modelling, qualitative spatial reasoning, place georeferencing, place knowledge querying
iv
[ April 29, 2019 at 20:02 – classicthesis version 5 ] DECLARATION
This is to certify that:
1. the thesis comprises only my original work,
2. acknowledgement has been made in the text to all other mate- rial used,
3. the thesis is less than 100 000 words in length, exclusive of tables and the bibliography.
Melbourne, Jan 2019
Hao Chen
[ April 29, 2019 at 20:02 – classicthesis version 5 ] [ April 29, 2019 at 20:02 – classicthesis version 5 ] PUBLICATIONS
This thesis is based on published works during the PhD research project. The major contents of this thesis such as ideas, algorithms, and figures have appeared previously in the following publications: journal articles
• H. Chen, M. Vasardani, S. Winter, and M. Tomko (2018). “A Graph Database Model for Knowledge Extracted from Place De- scriptions.” In: ISPRS International Journal of Geo-Information 7.6, p. 221
• H. Chen, M. Vasardani, and S. Winter (2018b). “Georeferencing places from collective human descriptions using place graphs.” In: Journal of Spatial Information Science 2018.17, pp. 31–62
• H. Chen, M. Vasardani, and S. Winter (2018a). “Clustering- based disambiguation of fine-grained place names from descrip- tions.” In: GeoInformatica peer-reviewed conference and workshop papers
• H. Chen, M. Vasardani, and S. Winter (2015). “Maintaining rela- tional consistency in a graph-based place database.” In: CEUR Workshop Proceedings - Locate15. Vol. 1323, pp. 1–12
• E. Hamzei, H. Chen, H. Hua, M. Vasardani, M. Tomko, and S. Winter (2018). “Deriving place graphs from spatial databases.” In: CEUR Workshop Proceedings - Locate18. Ed. by S. Peters and K. Khoshelham. Vol. 2087, pp. 25–32
vii
[ April 29, 2019 at 20:02 – classicthesis version 5 ] [ April 29, 2019 at 20:02 – classicthesis version 5 ] ACKNOWLEDGEMENTS
First of all, I wish to express my greatest gratitude to my supervi- sors Prof. Stephan Winter and Dr. Maria Vasardani for offering me the opportunity of doing PhD research, as well as for their continu- ous support throughout these years. I also want to thank them for their patience and encouragement. When I started with this project, I had great enthusiasm but little experience. They not only taught me the way of doing proper academic research, which is essential for becoming a competent researcher, but also provided me with advice and guidance when I experienced lost or frustration. I could not have imagined accomplishing this thesis or growing as a junior researcher without all their supports. I would like to thank my colleges: Yaoli, Hanxian, Haifeng, Ehsan, Ivan, David, Santa, Ha, Debaditya, Surabhi, Junchul, Michael, Marie, Rahul, Fuqiang, Zahra, Rajesh, Debjit, Subhra, Kourosh, Elham, and others, as well as people that I’ve worked with: Hua, Peng, Haonan, and Ping. It was fantastic to have the opportunity to work with you all. With a special mention to my committee chair Dr. Martin Tomko for his interesting, insightful comments that widen my research from various perspectives. I also thank him for accepting to be present each year for the review of my research progress. Besides, I want to thank the Australian Research Council for fund- ing the discovery research project DP170100109 that my PhD research sits within. This project also provides opportunities in the future for further extending the work of this PhD thesis. I am grateful to my parents, who have unconditionally provided me with economic, mental, as well as emotional support throughout my life in general, as well as during the time of my PhD research. I also want to thank them for encouraging me to think critically and to ask questions since my childhood. They helped me to develop my own interests and shaped my personalities. Without them I may not have chosen the path of academic research for seeking knowledge and fulfilling curiosity. Words cannot express how grateful I am to them for being great parents. Further, my sincere thanks goes to my friends: Won Chak, Chun, Xinying, Qihang, Yanjie, Jihong, Hanxian, Yaguang, and others, who supported and encouraged me in writing, and incented me to strive towards my goal. More importantly, they also reminded me that there is more to life than academic research. I wish to thank them for all the good times we spent together. Last but by no means least, my gratitude goes to Yundi, who has always been my support in the moments when I felt lost or down. I am grateful for her company, understanding, and thoughtfulness. Even though I have tried to balance between study and life, there was often not sufficient time and affordability for her from my side. She gave me great supports and immense encouragement so that I
ix
[ April 29, 2019 at 20:02 – classicthesis version 5 ] could accomplish my PhD research as well as this thesis. Thank you, yundi.
x
[ April 29, 2019 at 20:02 – classicthesis version 5 ] CONTENTS
List of Figures xiv List of Tables xx Acronyms xxi 1 introduction1 1.1 Background and Motivation ...... 1 1.2 Towards Building a Place-Based GIS from Natural Lan- guage Place Descriptions ...... 3 1.3 Tasks, Challenges, and Research Hypothesis ...... 5 1.4 Contributions and Significance ...... 9 1.5 Thesis Structure ...... 13 2 literature review 15 2.1 Place as a Cognitive Concept ...... 15 2.2 Place Models from an Information Systems Perspective 16 2.2.1 Gazetteer ...... 17 2.2.2 Place Semantics Models ...... 17 2.2.3 Modelling the Vagueness of Places ...... 18 2.2.4 Contrast Set of Places ...... 19 2.3 Spatial Language and Place Description ...... 20 2.3.1 Modelling Spatial Language from Text . . . . . 20 2.3.2 Natural Language Place Description ...... 21 2.4 Qualitative Spatial Relations ...... 23 2.4.1 Qualitative Spatial Representation and Reasoning 23 2.4.2 Modelling Qualitative Spatial Relations Quanti- tatively ...... 25 2.5 Georeferencing Place from Textual Documents . . . . . 27 2.5.1 Toponym Resolution ...... 27 2.5.2 Georeferencing Fine-Grained Places and Non- Gazetteered Place References ...... 29 2.6 Graph Representations of Spatial Knowledge ...... 31 2.7 Summary ...... 34 3 place knowledge modelling 35 3.1 Introduction ...... 35 3.2 Extending the Place Graph Model ...... 37 3.2.1 Information not Captured in the Basic Place Graph Model ...... 37 3.2.2 The Extended Place Graph Database Model . . 42 3.3 Implementation ...... 49 3.3.1 Data Overview and Graph Construction . . . . 49 3.3.2 Demonstrative example ...... 50 3.4 Chapter Discussion and Summary ...... 52 4 qualitative spatial reasoning 57 4.1 Introduction ...... 57 4.2 Consistency Reasoning in a Place Graph Database . . . 59 4.2.1 Problem Elucidation ...... 59 4.2.2 Single-Family Relational Inference ...... 61
xi
[ April 29, 2019 at 20:02 – classicthesis version 5 ] xii contents
4.2.3 Single-Family Relational Consistency Mainte- nance By Cycles ...... 67 4.2.4 Extending the Framework to Cross-Family Rea- soning ...... 69 4.3 Implementation and Experiments ...... 72 4.3.1 Single-Family Reasoning ...... 73 4.3.2 Cross-Family Reasoning ...... 77 4.4 Chapter Discussion and Summary ...... 78 5 place georeferencing 81 5.1 Introduction ...... 81 5.2 A Three-Step Georeferencing Approach ...... 83 5.2.1 Overview ...... 83 5.2.2 Step One: A Clustering Algorithm for Disam- biguating fine-grained place names ...... 84 5.2.3 Step Two: Deriving Approximate Location Re- gion Representations ...... 90 5.2.4 Step Three: Gazetteer Best Matching ...... 96 5.3 Implementation and Experiments ...... 99 5.3.1 Preprocessing ...... 99 5.3.2 Evaluation of DensityK ...... 100 5.3.3 Results and Discussions of the Three-Step Geo- referencing Approach ...... 109 5.3.4 Comparison with Georeferencing Using a Basic Place Graph ...... 115 5.4 Chapter Discussion and Summary ...... 119 6 place knowledge querying 121 6.1 Introduction ...... 121 6.2 Querying a Place Graph Database ...... 123 6.2.1 A Categorization of Place Knowledge Queries . 123 6.2.2 Query Framework ...... 126 6.3 Implementation and Experiments ...... 129 6.3.1 Query Demonstration ...... 129 6.3.2 Contextualized Querying ...... 132 6.4 Chapter Discussion and Summary ...... 135 7 place graph database management system 139 7.1 System Overview ...... 139 7.2 Database Management Module ...... 140 7.3 Input Preparation Module ...... 141 7.4 Database Construction Module ...... 143 7.5 Georeferencing Module ...... 143 7.6 Place Mapping Module ...... 145 7.7 Querying Module ...... 146 8 discussion and conclusions 147 8.1 Discussions of the Major Results ...... 148 8.1.1 Place Knowledge Modelling ...... 148 8.1.2 Spatial Relationship Reasoning ...... 149 8.1.3 Place Georeferencing ...... 151 8.1.4 Place Knowledge Querying ...... 152
[ April 29, 2019 at 20:02 – classicthesis version 5 ] contents xiii
8.2 Summary of Contributions and Evaluation against Hy- pothesis ...... 153 8.3 Limitations and Future Work ...... 156 8.3.1 Natural Language Processing for Place Knowl- edge Extraction ...... 156 8.3.2 Contextualized Spatial Relationship Modelling 157 8.3.3 Application of a place graph database ...... 158 bibliography 161
[ April 29, 2019 at 20:02 – classicthesis version 5 ] LISTOFFIGURES
Figure 1 An example of a short description about Fed- eration Square, a landmark in Melbourne, with several place names being mentioned (Source: http://www.travelandleisure.com)...... 5 Figure 2 Workflow showing the main tasks of this thesis and associated approaches...... 9 Figure 3 A simple place graph representing the spatial references “the courtyard is on the campus" and “the courtyard is beside the clocktower". . 32 Figure 4 An example basic place graph with node size corresponding to node degree, and edge size corresponding to number of relationships be- tween the linked nodes. Multiple relationships between two nodes are represented by only one edge...... 36 Figure 5 UML diagram of the basic place graph model. 38 Figure 6 UML diagram illustrating the extended place graph database model, with seven types of classes (nodes) and nine types of relationships (edges)...... 43 Figure 7 Examples of modelling a binary spatial rela- tionship west (left) and a non-binary spatial re- lationship between (right) using n-plet...... 43 Figure 8 An example of modelling a relative direction relationship using the extended place graph model...... 46 Figure 9 Creation of description and user nodes. . . . . 52 Figure 10 Creation of n-plet nodes...... 53 Figure 11 Creation of place reference, spatial relation, and route nodes...... 53 Figure 12 Linking different place reference nodes to the same place node through node merging. . . . . 53 Figure 13 The resulting extended place graph database of the place description example. Gray: de- scription; Pink: user; Green: n-plet; Blue: (mapped) spatial relation; Red: route; Yellow: place reference; Purple: place...... 54 Figure 14 An illustrative example of relational inference and consistency reasoning problems. The spa- tial configuration of three spatial objects (left), and two spatial relationships describing the configuration (right)...... 59
xiv
[ April 29, 2019 at 20:02 – classicthesis version 5 ] List of Figures xv
Figure 15 Qualitative spatial reasoning models of car- dinal direction relations proposed by Frank (1992), including a cone model (a), a half-plane model (b), and a neutral zone model (c). . . . . 61 Figure 16 A failure case of the proposed composition ta- ble for the neutral zone model when applied to natural language spatial relationship expres- sions...... 62 Figure 17 Refinement through ECD intersection with ad- ditional spatial relation knowledge...... 64 Figure 18 A qualitative distance relationship model for reasoning (left), and the limitation of the model when considering triangle inequality (right)...... 64 Figure 19 Consistency reasoning through reference di- rection transformation. An existing relation- ship (a); another new relationship that is con- sistent with the existing one, determined by reference direction transformation (b); another relationship that is inconsistent with the previ- ous two (c)...... 66 Figure 20 An example of consistency checking of a new triplet based on existing path and the derived existing knowledge...... 69 Figure 21 Consistency matrix, with green indicating that the two spatial relation are regarded as consis- tent with each other...... 70 Figure 22 The numbers and proportions of spatial rela- tions from each of the four spatial relation fam- ilies from the campus place graph database. . 72 Figure 23 Number of existing paths of cardinal direc- tion relationships queried by place id for all places in the campus place graph database, with (right) and without (left) considering in- verse relationships...... 74 Figure 24 Number of existing paths of relative direction relationships queried by place id for all places in the campus place graph database...... 74 Figure 25 Number of existing paths of qualitative dis- tance relationships queried by place id for all places in the campus place graph database. . . 75 Figure 26 Number of existing paths of topological rela- tionships queried by node id for all places in the campus place graph database...... 75 Figure 27 Query time (in seconds) for existing path with different lengths for individual places...... 76 Figure 28 The locations of the three places mentioned in the descriptions above, with a red arrow indi- cating the reference direction...... 76
[ April 29, 2019 at 20:02 – classicthesis version 5 ] xvi List of Figures
Figure 29 Number of existing paths of any spatial rela- tionships queried by place id for all places in the campus place graph database, with consid- ering inverse relationships (right) and without (left)...... 77 Figure 30 A sample representation of places linked by spatial relations. Each place is associated with one or more place references captured in a place graph...... 84 Figure 31 The workflow of this chapter, with the first three phases corresponding to the three major subtasks of the approach...... 84 Figure 32 Two example DensityK functions from dif- ferent input data with cluster distance high- lighted (a, b), and comparisons of DensityK functions generated based on annular and cir- cular search regions for the same data as in (a) and (b) respectively (c, d)...... 87 Figure 33 DensityK function generated with four dif- ferent ∆d intervals for the same input point cloud: 100, 250, 500 (meters)...... 88 Figure 34 Disambiguated anchor places a, c, d from the example shown in Figure 30, which are within the cluster identified as spatial context. (source: Google Map, 2015)...... 90 Figure 35 Search spaces for cardinal direction relations based on the centroid of the relatum (a, b, c), and an alternative model for non-point relata (d) that is not applied in this chapter...... 91 Figure 36 Search spaces for the qualitative distance rela- tions in this chapter (a, b, c), and a comparison to the model by Liu et al. (d)...... 91 Figure 37 Search spaces for relative directions given a reference direction...... 92 Figure 38 Search spaces for covered by, equal, inside (a), dis- joint, meet (b), and the other three topological relations overlap, cover, contain (c)...... 92 Figure 39 Example of the density surface generated by KDE (a), the density surface generated by re- gression (b), and the hexagon representation generated by tessellation (c, d) for near trained for a specified contextual criteria set, based on relative locatum locations (distance [m]). . . . 95 Figure 40 An example of deriving the ALR (the shaded region) for Place b through integrating three search spaces (Source: Google Maps, 2015) (a) and deriving an ALR by integrating two prob- abilistic search spaces into a new density sur- face (b) (distance [m])...... 96
[ April 29, 2019 at 20:02 – classicthesis version 5 ] List of Figures xvii
Figure 41 Illustration for spatial relation satisfaction for near with relatum in the middle (a), north of with relatum at the bottom (b), and overlap with relatum in the middle (c)...... 98 Figure 42 Numbers of ambiguous gazetteer entries of places names from the two datasets, campus (left) and Melbourne (right)...... 99 Figure 43 An example input point cloud of all ambigu- ous gazetteer entry locations of a set of place names from the campus dataset, with ground- truth locations highlighted in red color. . . . . 101 Figure 44 Clustering results generated by established clustering algorithms for place name disam- biguation...... 105 Figure 45 Clustering results generated by density-based clustering algorithms...... 106 Figure 46 Clustering results generated by hierarchical clustering algorithms...... 106 Figure 47 Clustering results generated by partitioning re- location clustering algorithms...... 107 Figure 48 Clustering results generated by other cluster- ing algorithms...... 108 Figure 49 Clustering results generated by the DensityK algorithm...... 109 Figure 50 Disambiguation precisions (left), and average distance errors in km (right) by individual doc- uments...... 109 Figure 51 Deriving cluster distance for input point clouds from the campus graph (top) and the Melbourne graph (bottom); generating clusters for disambiguation from point clouds (left), and disambiguated anchor places forming spa- tial contexts (right)...... 110 Figure 52 Comparison of KDE-based search space gener- ated by removing 80% of random input train- ing points (right) with the search space gener- ated with full training points (left) for testing the robustness of the training approach (dis- tance [m]) ...... 111 Figure 53 Search space examples for triplets with building-level locata that have certain spatial relationships to relata from different levels, with prominence and spatial discourse gran- ularity undetermined, generated by the KDE model (distance [m])...... 112
[ April 29, 2019 at 20:02 – classicthesis version 5 ] xviii List of Figures
Figure 54 Precision by best-matching scores (a); distance errors by place for the two graphs (b, c) for both formal and regression-based model (dis- tance [m]); precision and recall trade-off for identifying non-gazetteered places by thresh- olding (d)...... 114 Figure 55 Example of representing the location of swim- ming pool on map, given two spatial relation- ships to two anchor places. Search spaces of the two relations as contours (a); crisp ALR with 0.95 as threshold (b); crisp ALR with 0.5 as threshold (c); ground-truth location of the place (d) (Source: Google Maps, Jan 2018). . . . 115 Figure 56 The spatial context of a merged, original place graph (left), and separated spatial contexts of an extended place graph (right)...... 116 Figure 57 Search space of place B for relationship
[ April 29, 2019 at 20:02 – classicthesis version 5 ] List of Figures xix
Figure 70 Interface layout of the developed place graph database management system...... 140 Figure 71 Dependence relationships among the six func- tional modules and their main functions. . . . 140 Figure 72 Interface of the database management module in the module tab zone...... 141 Figure 73 A web-based game for collecting place descrip- tions...... 141 Figure 74 Workflow of generating place graphs from maps.142 Figure 75 Interface of the input preparation module in the module tab zone and its sub-tabs...... 142 Figure 76 Interface of the database construction module in the module tab zone...... 143 Figure 77 Interface of the georeferencing module in the module tab zone...... 144 Figure 78 Georeferencing results from each of the three steps shown in the textual output zone: dis- ambiguation of anchor places (top), deriving approximate location regions for non-anchor places (middle), and gazetteer best-matching of non-anchor places (bottom)...... 144 Figure 79 Interface of the place mapping module in the module tab zone (top), and an example of two places being mapped in the mapping zone with both their approximate location regionss and matched gazetteer entry locations (bottom). 145 Figure 80 Interface of the querying module in the mod- ule tab zone and its sub-tabs...... 146 Figure 81 An example of contextualized query through a dialog...... 146
[ April 29, 2019 at 20:02 – classicthesis version 5 ] LISTOFTABLES
Table 1 Properties of a place reference node...... 45 Table 2 Properties of an nplet node...... 45 Table 3 Properties of a place node...... 47 Table 4 Formal binary spatial relations from different families...... 47 Table 5 Properties of a spatial relation node...... 48 Table 6 Properties of a description node...... 48 Table 7 Frank’s composition table for the neutral zone model for cardinal direction relations Frank (1992)...... 62 Table 8 A modified composition table for cardinal di- rection relationships...... 63 Table 9 Extended cardinal direction relation sets of cardinal direction relations...... 63 Table 10 Modified composition table of qualitative dis- tance relations based on Algorithm 1...... 64 Table 11 Composition table of topological relations. . . 65 Table 12 Inverse relations of cardinal direction, relative direction, and topological relations...... 67 Table 13 Formal spatial relations and examples of their original natural language expressions from place descriptions...... 73 Table 14 Example of best-matching for node b based on computed overall scores...... 98 Table 15 Proportions of places from the three situations from the input datasets...... 100 Table 16 Parameter configurations of algorithms to be tested for place name disambiguation...... 103 Table 17 Average precision of each algorithm with the best-performing parameters on the tested datasets...... 104 Table 18 Precisions of anchor place disambiguation. . . 110 Table 19 Georeferencing performance by search space models for the two input graphs ...... 113 Table 20 Matching precisions with different weights of the overall measurement function when apply- ing the regression model ...... 115 Table 21 Types of spatial queries about spatial relation- ships...... 124 Table 22 Results for the three queries from the third ex- periment, with the second query using Alice Hoy as an example...... 131
xx
[ April 29, 2019 at 20:02 – classicthesis version 5 ] ACRONYMS
ALR approximate location region
API application programming interface
BFS breadth first search
CCS contextual criteria set
CRF conditional random fields
DFS depth first search
ECD extended cardinal direction relation set
EK existing knowledge
EP existing path
GIR geographic information retrieval
GIS geographic information system
GIScience geographic information science
GPR Gaussian process regressor
GUM GeneralizedUpperModel
HMM hidden Markov model
IE information extraction
KDE kernel density estimation
LSTM long short-term memory
NER named entity recognition
NL natural language
NLP natural language processing
OSM OpenStreetMap
PG place graph
PGD place graph database
POI point of interest
POS part-of-speech
QSR qualitative spatial reasoning
RDF resource description framework
xxi
[ April 29, 2019 at 20:02 – classicthesis version 5 ] xxii acronyms
RNN recurrent neural networks
SPG spatial property graph
SpRL spatial role labelling
TR toponym resolution
TGN Getty Thesaurus of Geographic Names
UML unified modeling language
[ April 29, 2019 at 20:02 – classicthesis version 5 ] INTRODUCTION 1
1.1 background and motivation
People talk about space by referring to places (Winter et al., 2010). The term place has existed for long in common language and the concept has also been studied extensively in philosophy, psychology, and geography (Relph, 1976). The notion of place is used to refer to individual meaningful portions of space and to describe the spatial locations of geographic objects. Place functions primarily as spatial anchor in cognitive and communicative tasks and has been regarded as the prototypical spatial reference in human, economic and culture geography (Cresswell, 2014). Compared to a space-based perspec- tive of geography, Tuan (1977) regards place as space infused with human meaning and experience thus enabling conversations. Place- based research is a research dimension with an increasing popularity in geographic information science (GIScience) in order to smooth and simplify human-computer interactions (Davies et al., 2008; Raubal, 2009; Winter et al., 2016). The objective is capturing the notion of place from cognitive and linguistic perspectives, thus enabling modelling and utilizing place-related information. In GIScience, the importance of place-based research has been widely acknowledged, (e.g., Golledge, 1997; Goodchild, 2007, 2011; Winter and Freksa, 2012; Winter et al., 2016). For instance, facing the emergence of ubiquitous computing with users no longer necessarily trained geographic information sys- tem (GIS) experts, Egenhofer and Mark (1995) suggested Naive Geog- raphy in order to capture and reflect the way that non-experts think and reason about space and time. Place is regarded as an elusive notion in GIScience. One reason is that places are often associated with inherently vague spatial meanings and is not immediately representable by the current GISs and spatial databases, while the latter ones are typically based on unambiguous, crisp, and metric geometries. This vagueness is evident in human cognition, perception, as well as natural language (NL) descriptions (Agarwal, 2005). Although the concept of place has been studied ex- tensively in GIScience, (e.g., Bennett and Agarwal, 2007; Davies et al., 2008; Vasardani, Winter, and Richter, 2013; Winter et al., 2010), there is currently no widely adopted computational data model allowing reasoning and inquiring of place knowledge from a human perspec- tive. Spatial cognitive representations seem to rely on places as enti- ties, and evidence has shown that these representations do not work on a coordinate-based geometry. Cognitive collage has been used as a metaphor for spatial mental models, which is often schematic or dis- torted compared to metric-based knowledge (Tversky, 1993). People mentally construct and talk about places through references to places (e.g., place names) and spatial relationships between places (e.g., near,
1
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 2 introduction
in front of ) that are typically qualitative instead of quantitative, often with a frame of reference. Such knowledge, despite often being vague, allows inference of spatial locations, although it also can be highly ambiguous if taken out of its conversational context. In contrast, GISs and spatial databases, as well as services built on them (e.g., webmap- ping, location-based services, and navigation services), typically rely on unambiguous, coordinate-based geometries with official names as the unique identifier to represent places. In short, place, while fun- damental to human cognition and communication, is still beyond the reach of current information systems. This fundamental mismatch has caused several obstacles in human- computer interactions about place, particularly when a computer at- tempts to interpret place references and spatial relations from spatial language (e.g., place descriptions and inquires). Regarding place, the same place reference could refer to different geographic locations or extents depending on the conversational context. This is partly due to the ambiguity of place references (e.g., Washington could refer to more than 40 populated places around the world 1, and train station as a place type is even more ambiguous), and partly due to the in- determinacy of place as a vague concept (e.g., when asked, people draw different boundaries for the place reference downtown (Taylor et al., 2009), as individuals think about place boundaries imprecisely and differently). Thus, it seems inappropriate to represent places by unambiguous geometry with official names as the identifier. Regard- ing spatial relationship, it is also challenging for current systems to interpret spatial relationships from spatial language, as well as to per- form reasoning with them. For example, webmapping systems when queried for a place typically ignore the semantics of spatial relations such as in front of and at, and return results based on optimized func- tions considering string matching, distance, and relevance. This way of responding is different from how a human would interpret a spa- tial inquiry and give answers accordingly, and thus often lead to un- wanted or incorrect results. Some failure examples are provided (Win- ter and Truelove, 2013). Spatial language that requires reasoning and inferences in order to decode, e.g., “inside the square that is to the south of the University of Melbourne”, is even more challenging and beyond the capacity of these systems. These obstacles have caused frustrations of people when commu- nicating with information systems such as for emergency services, web searches, trip planners, or navigation (Vasardani, Winter, and Richter, 2013). If everybody were satisfied with such information sys- tems, then place modelling would not be an issue (Davies et al., 2008; Winter and Truelove, 2013). The essential cause is the incompatibility between the cognitive representation of place and their spatial rela- tionships (as is reflected by how people communicate about places) and the design of current GISs and spatial databases, which hinders these systems from reasoning and communicating about place knowl- edge from a human perspective. To overcome these obstacles, the core
1 https://en.wikipedia.org/wiki/Washington
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 1.2 towards building a place-based gis from natural language place descriptions 3
challenge is to design an information system from a different end that focuses on place knowledge modelling from a human perspec- tive. The difficulty comes from the variety of spatial representations and semantics that individuals as unique sensors assign to places. Consequently, place modelling may seem more natural to be done in a collective, generalized manner. Since such knowledge is typically context-dependent instead of absolute, it is important to characterize and capture context in such a system as well. With such a system as a place knowledge base built from a human perspective, it would be possible to further explore how this system could be used for improving current services. From a language com- prehension perspective, the ability to understand human spatial de- scriptions and linking to locations in space appears to be useful for systems such as emergency response and autonomous vehicles. From a querying perspective, Ed Parsons, Geospatial Chief Technologist of Google, indicated that “about 1 in 3 of queries that people just type into a standard Google search bar are about places” 2, and thus, there is a strong motivation to improve place query mechanisms and pro- vide more relevant results. From a description generation perspective, NL is preferable in many situations than other representations of spa- tial knowledge such as maps, which are often unintuitive, contain irrelevant details to the task at hand, and mentally costly to interpret. The ultimate goal of such a system is to be used for building the spatially intelligent machine proposed by Winter and Wu (2009) that is capable of understanding and producing human spatial language, thus allowing smooth communication between human and machine.
1.2 towards building a place-based gis from natural language place descriptions
This PhD research has been developed towards the goal of building a place-based GIS (Elwood, Goodchild, and Sui, 2013). NL place descrip- tion is regarded as the target source of this thesis, as they reflect the way that people mentally represent and communicate about place knowledge. This thesis aims at contributing to several research as- pects in order to identify and minimize gaps in knowledge, specif- ically modelling, processing, and utilizing of place knowledge ex- tracted from place descriptions. Therefore, this thesis involves connec- tions between several disciplines such as language, spatial cognition, and information systems. In terms of information extraction (IE), this thesis will use off-the-shelf natural language processing (NLP) tech- niques instead of developing new ones. The outcome of this thesis is twofold. The first outcome includes novel developed approaches ad- dressing these aspects as scientific contributions. The second outcome is an implemented system that captures information about places that place descriptions contain, as well as allows processing of the captured information in order to use them in application scenarios (e.g., mapping or querying). The system is able to provide selective
2 http://youtu.be/ucYiMBfyNfo
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 4 introduction
information about places that the descriptors (people giving the de- scriptions) found worth describing. Such information is collective and could be complementary to the information that traditional GISs and authoritative databases provide. Place descriptions occur in everyday verbal communication as a way of encoding and transmitting knowledge about places between individuals (Vasardani et al., 2013). The information conveyed is use- ful for mental-sketching of a spatial environment, and can be used to, for example, provide navigational instructions or inform the loca- tion of events. They typically provide qualitative reference systems for describing the locations of places using place references and spa- tial relations, together with other non-spatial information, such as human activities and place characteristics. An example description from a real-world dataset is shown below, which communicates in- formation about the University of Melbourne campus environment by showing several places and their mutual spatial relationships. It also describes the characteristics and the functions of some of the places. As a new type of data source that is unstructured and unsup- ported by current GISs and spatial databases, it remains challenging to develop methodologies for modelling, processing, and utilizing the knowledge contained by such place descriptions.
“. . . South Lawn is the major reference point which is situ- ated in about the middle of the campus. Coming from the Main South Entry, the Baillieu Library will be on the left- hand side of the South Lawn. To the north of this you have the Old Quad (really old English style building). If you want food and are currently on South Lawn go through the Old Quad to the north and keeping heading north un- til you get to a Union House ...”
The web provides a plethora of place descriptions in forms such as news articles, social media texts, trip guides, and tourism articles as a rich source for constructing a place information system. Figure 1 presents a place description as web document from a tourism website. Such a document can be collected through Web-harvesting techniques (Kim, Vasardani, and Winter, 2015). Web-based place descriptions of- fer a unique opportunity for the generation of rich datasets about human place knowledge. This thesis aims to construct a place knowledge base from natural language place descriptions and utilize the captured knowledge in different ways. The structure of place descriptions has been studied extensively in linguistics and spatial cognition research (e.g., Coucle- lis et al., 1987; Geeraerts and Cuyckens, 2007; Jarvella and Klein, 1982). Recently, the relation between place and language has attracted re- search attention in the domain of GIScience as well, addressing issues related to language, geography, and cognition. The driving force is the increasing volume of unstructured text documents being pub- lished online (Melo and Martins, 2017), as well as the growing need for place-related information in everyday life. Place is also an im- portant aspect in social media-related research, as a wide range of
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 1.3 tasks, challenges, and research hypothesis 5
Figure 1: An example of a short description about Federation Square, a land- mark in Melbourne, with several place names being mentioned (Source: http://www.travelandleisure.com). social media platforms structure users’ activities by places (possibly also with geotagged locations), such as Facebook, Twitter, Foursquare, Flickr, and Instagram. It has been argued that users might greatly benefit themselves by having richer and more locally valid data on place in many situations with place information extracted from tex- tual documents online (Davies et al., 2008). Thanks to the rapid devel- opment of text mining and NLP techniques, parsers have already been developed to extract multiple types of information from text docu- ments including place descriptions, such as place references (e.g., De- lozier, Baldridge, and London, 2015; Lieberman, Samet, and Sankara- narayanan, 2007; Moncla et al., 2014) and spatial relationships (e.g., Khan, Vasardani, and Winter, 2013; Kordjamshidi, Van Otterlo, and Moens, 2011; Liu, Vasardani, and Baldwin, 2014).
1.3 tasks, challenges, and research hypothesis
According to the objective described above, this thesis identifies four major tasks, namely place knowledge modelling, qualitative spatial reasoning (QSR), place georeferencing, and place knowledge query- ing. There are several challenges associated with each of these tasks, which are discussed below in this section.
Place Knowledge Modelling
Place knowledge is relational instead of absolute, and Agnew (2011) suggests thinking of place in relation to other places, instead of iso-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 6 introduction
lated features. Place descriptions typically consist of place references linked by spatial relations, called locative expressions and naturally con- tain relational knowledge between places. Locative expressions have been suggested to be modelled in the form of triplets (Vasardani et al., 2013) of a locatum L, the reference to a place that is to be located, a relatum R, the reference to a place that is already located, and a spa- tial relationship r between the two, e.g.,
Qualitative Spatial Reasoning
As a place knowledge base with qualitative spatial relationships cap- tured, it should ideally be used for QSR tasks in a way similar to a human. Specifically, it should be able to conduct inference – deriving new relationships based on existing relationships, as well as consis- tency reasoning – identifying logical contradictions among the stored relationships, rather than accepting all inputs as they are. For exam- ple, with two relationships (here as triplets) between three places
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 1.3 tasks, challenges, and research hypothesis 7 the new knowledge. People may make mistakes when giving place descriptions, and malicious users could provide inconsistent knowl- edge. Therefore, the capacity of QSR is preferable for maintaining the consistency in such a knowledge base. A human is capable of such reasoning tasks based on mental mod- els (Knauff et al., 1998), at least for simple scenarios with few places and spatial relationships. In contrast, the system to be developed may form a constrained relational network that applies QSR rules (e.g., Cohn and Renz, 2008; Wallgrün et al., 2008). QSR is a well-developed research field with formal rules defined for multiple spatial relation families. However, these models have not been used for reasoning with collective human spatial knowledge before, and they may not be flexible enough to accommodate NL spatial relationship expres- sions. Moreover, a mechanism to apply QSR models from multiple spatial relation families for consistency maintenance considering the data structure of the system has to be developed.
Place Georeferencing
Establishing links between the stored place knowledge in the system as well as other GISs, spatial databases, and spatial services could increase the usefulness of the system through providing complemen- tary knowledge. For many application scenarios, such as mapping and location-based analysis, establishing the link would require the places in the system to be located in space through georeferencing. Various approaches have been developed for georeferencing place from textual documents, and a typical approach, called toponym res- olution (TR) (Leidner, 2007), relies on place name gazetteers match- ing. These approaches typically focus on larger geographic features such as populated places (e.g., cities or towns) or natural geographic features (e.g., rivers or mountains). For these features, disambigua- tion heuristics can leverage the size, population, or containment rela- tionships of candidate places, possibly based on external knowledge bases. However, place descriptions occur in everyday conversations often contain references to fine-grained places, e.g., streets, buildings, and local point of interests (POIs), which are often significantly more numerous and more similar thus require a different method to re- solve. Moreover, place descriptions are flexible, vernacular, and of- ten contain place references that cannot be found in gazetteers, such as synonyms or place types (e.g., the large square). Conventional ap- proaches quickly fail when facing such references. In addition, place references may refer to places that are not gazetteered at all, such as places from environments that are too fine-grained or conversational contexts that are too limited to be captured by gazetteers (e.g., the dean’s office). Spatial relationships in place descriptions provide useful informa- tion for place localization. In a triplet, the locatum is located through providing a spatial relation to the relatum, which is typically a land- mark whose location is better known, and thus, in relation to which
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 8 introduction
the location of the locatum can be specified. Yet such a relationship does not explicitly provide any locative information of the locatum, and thus, cannot be immediately interpreted by a computer, due to the vagueness of spatial relationships (Fogliaroni and Hobel, 2015; Liu et al., 2009b). Consequently, although it is usually easy for a hu- man recipient of a description to disambiguate place references or approximately locate places by spatial relationships, computationally modelling spatial relationships from NL to allow automatic interpre- tation remains a significant and open challenge. Through leveraging spatial relationship knowledge among places once it is modelled, non-gazetteered references to places that cannot be located previ- ously may be georeferenced, thus higher georeferencing recall can be achieved. Meanwhile, the semantics of spatial relationships provides potentially useful knowledge for place reference disambiguation in certain conversational contexts, and therefore, higher georeferencing precision is expected once spatial relationships are considered in the georeferencing process.
Place Knowledge Querying
As a knowledge base with collective information about places that people found worth describing, the system to be developed should also support querying of the stored knowledge. Such knowledge can then be utilized in future application scenarios. Therefore, place knowledge query is selected as the final task of this thesis. The corresponding subtasks include identifying types of queries that can be answered by the knowledge captured in the system, as well as developing query mechanisms. Another challenge to be addressed in this thesis is querying by context. Knowledge originally extracted from descriptions with different conversational contexts may have different degrees of relevance regarding a query, if the knowledge is context-dependent. Conventional queries by exact-matching will return results aggregated from possibly multiple sources of place de- scriptions with various contexts, which may be less useful compared to results already filtered by contexts. For example, it is perfectly possible to collect seemly contradicting information from different place descriptions, which are all true in their own conversational contexts. Consequently, it is necessary to develop a contextualized query approach for the system to be developed. The query approach should have some mechanisms to be able to differentiate knowledge from different contexts and return only knowledge that is most relevant to the context given a query.
According to the identified major tasks and challenges, the hy- potheses of this thesis include:
• A place graph database (PGD) model can be developed for cap- turing place-related and contextual information extracted from place descriptions, and the model could overcome limitations
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 1.4 contributions and significance 9
of the previously-proposed place graph (PG) model and achieve better results in reasoning, georeferencing, and querying tasks.
• Spatial relational consistency can be preserved after transac- tions (i.e., creation, update, deletion) happen within a PGD through relational inference and consistency reasoning, and the place database will always end up in a consistent state.
• The PGD model allows achieving higher precision and re- call than typical TR approaches when georeferencing places from collective place descriptions, through considering non- gazetteered references and semantics of spatial relationships.
• The PGD model allows answering different types of queries including ones that cannot be answered by the previously- proposed PG model. Moreover, query results can be contextu- alized based on the captured context information.
1.4 contributions and significance
This thesis contributes towards modelling, processing, and utilizing human place knowledge extracted from NL place descriptions. The major outcome of this thesis is an implemented PGD, with scientific contributions include approaches proposed for the identified four ma- jor tasks from the last section. The workflow of this thesis is shown in Figure 2.
Figure 2: Workflow showing the main tasks of this thesis and associated approaches.
The four tasks are sequentially linked instead of isolated, and thus are addressed by four phases in this thesis.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 10 introduction
Modelling
The first phase focuses on developing a graph database model for cap- turing information extracted from place descriptions. Different types of information that are embedded in place descriptions, including contextual information, are identified in the first place. There are dif- ferent types of contexts, and this thesis adopt the categorization of description-, environment-, and human-dependent ones Wolter and Yousaf (2018). Then, a PGD model is designed in a way that allows con- venient and efficient query of the modelled information for the later three phases. Different types of information are captured by nodes and edges with different labels (denoting node types) as well as prop- erties (as node and edge attributes). The implemented module in this phase enables creating a PGD with a structured JSON with parsed in- formation from place descriptions as input. The module also contains an interface for visualizing and interacting with the database. For evaluation, the performances of the model are compared to the most similar and competitive model by Vasardani et al. (2013) in the later chapters, and the superiority of the model is confirmed. Al- though the model is designed specifically for place descriptions and for the three tasks (reasoning, georeferencing, and querying), discus- sions comparing the model to other spatial language models are pro- vided as well. The model allows extensions considering future needs for modelling other elements from spatial language. For example, al- though locomotions such as walk and go, which are common in route descriptions (Winter et al., 2018), are currently not an element in the model, they can easily be modelled by specifying a new node type in a way similar to spatial relationships.
Reasoning
In the second phase, spatial relationships captured in a PGD are used for relational inference and consistency maintenance. The reason for introducing this phase is that spatial relationship knowledge is essen- tial for both locating places (in the third phase) and querying place knowledge (in the fourth phase), and therefore, it is useful to de- rive additional spatial relationships and to ensure there is no logi- cal contradiction among the stored relationships in a PGD. For these purposes, relational composition rules for four spatial relation fam- ilies are defined. Specifically, reference frames, as one type of the contextual information captured, is used for reasoning with relative direction relations extracted from different contexts. Then, a path con- sistency checking algorithm is conducted for validating local consis- tency. Paths of different limited lengths are tested and compared, con- sidering that human spatial knowledge is typically local (e.g., Warren et al., 2017), and compositing longer paths are computational costly and likely to result in universal relationships anyway. The module developed in this phase will flag inconsistent relationships detected given a PGD as input. It will not decide which relationship is true if
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 1.4 contributions and significance 11 there is a contradiction, however, since the decision cannot be made without ground-truth knowledge. The proposed approach is a first-step to demonstrate the feasibil- ity of reasoning with human spatial knowledge about places using a PGD with QSR models. Currently four types of spatial relation families are considered, since they are relatively well-studied in QSR. Other re- lations, including non-binary ones such as between and through, are more challenging to be applied for reasoning. Contextual informa- tion has also been leveraged for contextualized reasoning, since the same spatial relation may have different semantics in different con- texts, and therefore may apply different reasoning rules. In this thesis one type of context (reference direction) is considered. The reason- ing framework proposed allows future extensions of new spatial rela- tions and context factors through adding associated reasoning rules. Finally, this thesis provides insights of cross-family reasoning for fu- ture research based on observations from preliminary experiments.
Georeferencing
In the third phase, places in a PGD is georeferenced in order to be located in space. Georeferencing enables mapping of these places, as well as to establish links between the database and other GISs and spatial databases for applications. The challenges of this task have been discussed, and this phase is divided into three steps correspond- ing to three sub-challenges. In the first step, places with gazetteered place references, as the easiest to resolve, are identified, and a density- based clustering algorithm is developed for disambiguating and lo- cating these places. The second step derives approximate location representations for each of the remaining places that does not have any gazetteered references, based on their spatial relationships to the places resolved in the first step. The approximate location representa- tions are derived based on both formal spatial relation search space models as well as probabilistic search space models trained from data. Finally, the approximate location representations are used for match- ing to gazetteer entries, based on a comprehensive similarity mea- suring algorithm considering string, semantic, and spatial similarity. Even if the matching fails, i.e., the place is non-gazetteered, the pre- viously derived approximate location representation can still be used to visually locate this place on a map. Therefore, the contributions of this phase is threefold aiming at resolving places from different cases, considering the mapping be- tween place and place reference. Accordingly, the georeferencing result can either be an entry from a gazetteer with a geometrical represen- tation (typically a point) for places from the first and second cases, or a density surface for places from the third case. Density surface- based representations are becoming increasing popular to be applied for places with indeterminate locations or boundaries (e.g., Gao et al., 2017a; Jones et al., 2008a), which are typically derived from collec- tive location data (e.g., geotagged images or texts from social media).
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 12 introduction
In comparison, this thesis shows how the locations of places can be derived as density surfaces based on collective spatial relationship knowledge with search space models.
Querying
Lastly, the fourth phase focuses on identifying types of queries that can be answered by the knowledge captured in a PGD, as well as devel- oping a query framework that can be applied to these types of queries. Since the developed database is graph-based, the query framework is based on graph traversal operations. Graph traversal is a typical way of querying a graph database, which matches nodes, edges, as well as paths by their types and properties according to the criteria given by a query (e.g., finding place nodes that have any spatial relationships to a specified place node). The developed query framework accepts both graph traversal queries as well as structured queries as input, and structured queries will be translated into graph traversal queries in order to be conducted. Moreover, the framework also leverages the captured contextual information in a PGD for filtering the stored knowledge according to contexts, based on a developed dialog-based query interface. Given a query, the system first retrieves candidate results from a PGD through conducting a graph traversal operation. Then, it analyzes these candidate results and group them by contexts. Finally, a dialog is initiated for letting users choose from the identi- fied contexts, and then the system will return filtered results accord- ingly. The query framework is demonstrated using different types of queries, and it is then evaluated through comparing to query results that are generated without the contextualization process. As mentioned previously, the ultimate goal of a place-based infor- mation system is a spatially intelligent machine that is capable of understanding and communicating place knowledge smoothly with a human. While the previous three phases focus on information mod- elling and processing, which can be regarded as understanding, query- ing is one of the core parts of communicating as it retrieves knowledge stored in such a system in order to use it in application scenarios. The next step of communicating is producing human understandable spa- tial language from the queried knowledge, which is out of the scope of this thesis.
Summary
The major contributions of this thesis are summarized as follows.
• A designed PGD model for efficient storing and retrieving of place information as well as contextual information extracted from place descriptions. The superiority has been demonstrated by comparing the performances to a competitive model in dif- ferent tasks;
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 1.5 thesis structure 13
• A reasoning framework for spatial relationship inference and maintaining local relational consistency in transactions hap- pened within a PGD. The framework is defined over four spatial relation families leveraging relational composition, path knowl- edge, and consistency rules;
• A multi-step georeferencing approach for locating all the stored places in a PGD regardless of whether they are referred by gazetteered references, with sub-contributions including: a novel clustering algorithm for place disambiguation; formal and probabilistic-based spatial relationship search space models; an approach to derive a density surface from search spaces to repre- sent the approximate location of a place; and a weighted multi- value gazetteer matching approach;
• A graph traversal-based query framework that accepts multiple input formats for different types of queries, including a dialog- based mechanism for contextualizing query results through identifying contexts, requesting additional user input, and fil- tering from candidate results; and
• An implemented PGD management system that integrates the above approaches into a complete processing chain, together with several other functionalities such as database visualization and place mapping.
1.5 thesis structure
The remaining chapters of this thesis are organized as follows. Chapter 2 provides an overview of the previous works that are related to place, place modelling, spatial language, qualitative spatial reasoning, georeferencing place from textual documents, and graph representations of spatial knowledge. Chapter 3 shows the proposed structure of the PGD with all the identified types of information from place descriptions being mod- eled. Each type of the identified information is discussed first. Then, the PGD model is introduced, and different types of nodes, edges, as well as their associated properties in the model are explained. The chapter also introduces the experimental datasets that will be used in the later tasks. This chapter addresses the first hypothesis. Chapter 4 first describes how spatial relations from different fami- lies can be composited for inferring new relationships. Then, a path- based consistency maintenance algorithm is introduced. The exper- iments focus on evaluating the approach and testing with different path length limitations. This chapter addresses the second hypothe- sis. Chapter 5 illustrates the three-step approach for georeferencing all places in a PGD, regardless of whether they have been referred by gazetteered place names or not. Places in different situations are eval- uated separately in the experiments with evaluation metrics includ-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 14 introduction
ing precision and distance error. This chapter addresses the third hy- pothesis. Chapter 6 first provides a categorization of types of queries that a PGD is capable of answering, regarding knowledge modelled by nodes, edges, paths, and associated properties. Then, the developed query framework is introduced. The framework is demonstrated with example queries and then evaluated. This chapter addresses the last (fourth) hypothesis. Chapter 7 provides a system description of the developed PGD man- agement system that integrates the approaches developed in the pre- vious chapters. Its functionalities are demonstrated with examples using the experimental datasets. Reviewing the proposed approaches and their experimental results, Chapter 8 presents a detailed discussion and relates the findings back to the hypotheses. The chapter then concludes with a summary of the contributions of this thesis, as well as suggestions for future work.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] LITERATUREREVIEW 2
According to the major challenges identified in the previous chapter, this literature review outlines the background with relevant work and highlights their relationships to this thesis. Section 2.1 commences with a review of how place, as a cognitive concept, is understood and defined in the literature. Next, existing models for place from an information system perspective are examined in Section 2.2. Sec- tion 2.3 introduces previously proposed annotation frameworks for decoding spatial-related language, as well as studies about the struc- ture of place descriptions. Then, Section 2.4 provides an overview of previous work about modelling qualitative spatial relationships, in particular from qualitative reasoning and quantification perspectives. Section 2.5 surveys research with regards to georeferencing places from text. Finally, Section 2.6 explains the concepts of spatial prop- erty graph (SPG), knowledge graph, and graph database. Section 2.7 summarizes the knowledge gap in the existing literature that is being addressed by this thesis.
2.1 place as a cognitive concept
The word place has existed for thousands of years in pragmatic com- mon language, and the concept has been studied in transdisciplinary research including philosophy, psychology, and geography. It is a ba- sic notion in everyday communication and plays a key role in almost every field of human enquiry (Canter, 1977; Harrison and Dourish, 1996; Jordan et al., 1998). Places, as cognitive concepts, emerged in a number of ways, when they are named by people, when people tell stories about them, and when people are doing things there (Agar- wal, 2005). Evidence from cognitive science and biological literature also shows that place is associated with ascertaining a ‘memory of a location’ (McNamara, 1991), and with the distance and direction estimates for locations in the environment (Chown, Kaplan, and Ko- rtenkamp, 1995). The concept of place is often compared to space, as the two are both fundamental concepts in geography, and more broadly in social sci- ences, humanities, and information science. People talk about space by referring to places (Winter et al., 2010). Place is regarded as a cen- tral concept in spatial cognition and communication, closely related to human experience (Couclelis, 1992; Tuan, 1977). In other words, it is human’s social interactions and experiences that turn space into places, infuse places with meanings, and enable conversations. Both space and place are linked to the ten core concepts (i.e., location, field, object, network, and event in the latest form) of spatial information pro- posed Kuhn (2012).
15
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 16 literature review
Recently in the discipline of geographic information science (GIScience), place-based research is receiving an increasing amount of attention, with a strong motivation in smoothing human-machine in- teractions. The major goals are capturing and modelling the concept of place from a cognitive perspective, as well as utilizing place-related knowledge in information systems. Extensive research have already stressed the importance of place-based research (e.g., Golledge, 1997; Goodchild, 2007, 2011; Winter and Freksa, 2012; Winter et al., 2016). However, place-based research is still relatively under-researched and less mentioned among GIScience topics, despite the recognized impor- tance (Winter, Kuhn, and Krüger, 2009). Davies et al. (2008) argues that the concept of place is, at best, an uncomfortable challenge for GIScience. Places tend to be almost defined as undefinable in the lit- erature, both spatially and semantically, with little agreement on its nature. Places are often spatially vague. The vagueness is evident in human cognition, perception, as well as in natural language (NL) communica- tions (Agarwal, 2005). The vagueness has also made place incompat- ible with current geographic information systems (GISs) and spatial databases that are based on crisp, unambiguous geometries removed from human concepts (Vasardani, Winter, and Richter, 2013). In cur- rent GISs, spatial databases, as well as services built on them, place is usually defined by textual place names associated with coordinate locations without further considering people’s perception and cog- nition factors. In comparison, it has been argued that places do not have any natural boundaries, and are locations that have been given shape and form by people (Winter and Freksa, 2012). Agnew (2011) also suggests thinking of place in relation to other places, instead of bounded and isolated features. As a result, currently there is still not much success in modelling and utilizing place knowledge (Adams and McKenzie, 2013). Some researchers further argue the concept of place may be too vague to be formalized, except in narrow circum- stances (Goodchild, 2011). Other research has also investigated what features make a place (Vasardani and Winter, 2016), while the diffi- culties to identify such features are acknowledged (Agarwal, 2005). In short, place, while fundamental to human cognition and commu- nication, is still well beyond the reach of current information systems.
2.2 place models from an information systems perspec- tive
It has been argued that existing GISs and spatial databases are mature in representing space, but limited in representing place (Gao et al., 2017b). However, integrating a model of how people conceptualize and perceive places into GISs is likely to increase the usefulness of these systems in many situations that requires human-computer in- teractions. In this section, a review of previous work on attempting to capture the notion of place from an information system perspective is provided.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 2.2 place models from an information systems perspective 17
2.2.1 Gazetteer
A gazetteer is often regarded as a geospatial dictionary of geographic names and typically contains three core components: place names, feature types, and footprints (Goodchild and Hill, 2008; Hill, 2000). A place name is what people usually search for this place, and is usually considered as the official name of the place. A place type is a category from a feature-type thesaurus for classifying places ac- cording to their semantics. A footprint represents the location of a place, typically by a single coordinate pair as an estimated center of an extended object, and sometimes by a polygon or a polyline in- stead. Some gazetteers, such as the Getty Thesaurus of Geographic Names (TGN)1 or GeoNames2, also store alternative names, and pro- vide detailed descriptions of places as well as positions of places in administrative or political hierarchies. Gazetteers are an important component in geo-referencing systems for both enterprise and aca- demic purposes, and are commonly used for geographic information retrieval (GIR) (Jones and Purves, 2007; Purves et al., 2007; Silva et al., 2006), navigation services and web-mapping applications.
2.2.2 Place Semantics Models
Existing GISs and spatial databases are limited in representing the se- mantics of places. A major challenge comes from the variety of seman- tics individuals assign to places, and it is difficult to define a model that captures such variety of place semantics in an information sys- tem. The way that a gazetteer maps space into places, though useful, does not always match the way people think about their world. Peo- ple assign complex meaning structures to places and decide their ac- tions and behavior based on such meaning. However, current GISs and spatial databases do not easily allow mappings of these meanings. With the rise of the concept of Semantic Web (Berners-Lee, Hendler, and Lassila, 2001), techniques for building a geospatial semantic web (Egenhofer, 2002; Fonseca, 2008; Janowicz et al., 2010) have seen a rapid development. One research direction is to model the seman- tics behind place names in gazetteers. Studies were made to enrich gazetteers considering ontologies of formal place semantics (Janowicz and Keßler, 2008; Janowicz, 2006). It is argued that such place repre- sentations with enriched semantics promises improvements in tasks including GIR, query expansion (Fu, Jones, and Abdelmoty, 2005a; Jones et al., 2004), and gazetteer conflation (Hastings, 2008; Keßler, Janowicz, and Bishr, 2009; McKenzie, Janowicz, and Adams, 2014). The feasibility of constructing a gazetteer in a bottom-up manner has also been demonstrated using semi-structured web documents and cloud computing techniques (Gao et al., 2017b). Bennett and Agarwal (2007) argue that a clear account of the nature of place can be elicited from an analysis of the semantic content of
1 http://www.getty.edu/research/tools/vocabularies/tgn/ 2 http://www.geonames.org/
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 18 literature review
NL. Therefore, it seems also promising to extract place semantics, e.g., based on place type keywords, events, and human activities, from unstructured, NL-based texts. Alazzawi, Abdelmoty, and Jones (2012) propose an automatic approach for the identification and extraction of service and activity-related concepts from textual place descrip- tions with the help of natural language processing (NLP) techniques, specifically lexical analysis learning tools. There are also other stud- ies that leverage language models for semantically describing places through word embeddings (Adams and Janowicz, 2015; Eisenstein et al., 2010; Speriosu et al., 2010). For example, Adams and Janowicz (2015) apply Latent Dirichlet allocation (Blei, Ng, and Jordan, 2003), a popular unsupervised topic learning technique, on Wikipedia arti- cles and generate thematic signatures for places by a mixture of topics with values. The signatures can be used to distinguish, (hierarchically) cluster, and measure similarity among places. People interact with places in their everyday lives, and places pro- vide a context for everyday actions. The affordance theory (Gibson, 1979) investigates how people perceive their environment, as well as what objects or things offer people to do with them. Regarding places, affordance refers to those properties of a place that determine certain human activities. It has been argued that place affordance is a core component for defining place, and it is therefore necessary to con- sider affordance when designing ontologies for places (Frank, 2003; Kuhn, 2001). Place type categories used in gazetteers can be regarded as taxonomies classifying place affordances, although sometimes lack flexibility and interoperability. Jordan et al. (1998) first proposed an affordance model for describing place, with the starting point that a place has to be an integration of both its location and its meaning in the context of human action. Later, more research attempted to study and formalize affordance in the domain of GIScience, although not nec- essarily all for places (Kuhn, 2007; Raubal and Kuhn, 2004; Scheider and Kuhn, 2010; Sen, 2008). The general underlying assumption is that affordance-oriented place ontologies are helpful to support the increasingly more complex applications requiring semantically richer conceptualization of the environment.
2.2.3 Modelling the Vagueness of Places
Capturing the uncertainties associated with places is one of the major challenges of place-based research. Vague place references such as The Midwest, or the The Alps are often used by people and have no crisp natural boundaries for defining their spatial extents (Hollenstein and Purves, 2010; Schockaert, 2011). Some models have been suggested to cope with such uncertainty, such the ’egg-yolk’ model (Cohn and Gotts, 1996), fuzzy sets (Fisher, 1996), rough sets (Bittner and Stell, 2000), broad boundary regions (Clementini and Felice, 1996), and su- pervaluation (Kulik, 2001). However, none of these approaches have been widely adopted, and they are argued to be inadequate due to the fact that they focus on the boundary rather than on the central no-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 2.2 place models from an information systems perspective 19 tion of the place (Winter and Freksa, 2012). The same place reference can refer to different spatial extents in different contexts. For exam- ple, the reference Federation Square in Melbourne, Australia can refer to either a set of connected places at the banks of the Yarra River, or the limited open area at the centre of these places. Other studies focus on capturing the uncertainty of the locations of places instead of their boundaries. Methods were developed to combine locality uncertainty measurements from different sources, such as imprecise distance, un- known datum, and map scale (Guo, Liu, and Wieczorek, 2008; Wiec- zorek, Guo, and Hijmans, 2004). Probabilistic and membership-based approaches, recently with the help of web-harvesting techniques, have been used to visually rep- resent the footprint of vague places. These approaches typically rely on aggregation of point locations provided by people that are consid- ered to be corresponding to a place. Such representations visualize the degrees to which any location belongs to such a place. Taylor et al. (2009) conducted a study to determine the footprint of down- town Santa Barbara by asking participants to draw the boundary of it and aggregating the results using density shading, as individuals think about place boundaries imprecisely and differently. Later, data- driven methods were proposed using techniques such as kernel den- sity estimation (KDE) and clustering, based on geotagged social media content of place names or tags, e.g., from websites such as Flicker or Instagram (Gao et al., 2017a; Grothe and Schaab, 2009; Hollenstein and Purves, 2010; Jones et al., 2008a). Other studies derive continu- ous fields of places with certain thematic topics (e.g., forests) over earth surface, based on interpolating from sources such as Wikipedia articles (Adams and McKenzie, 2013). There are also other function- based models that are typically based on web-harvested data for iden- tifying places associated with functional semantics such as commer- cial or green areas (e.g., Papadakis et al., 2019; Weerdenburg et al., 2019). These representations introduced above computationally char- acterize the spatial vagueness of places, and enable approximate crisp boundaries to be derived for place-based applications.
2.2.4 Contrast Set of Places
Winter and Freksa (2012) suggest to capture the cognitive and linguis- tic nature of a place in contrast to other places that are relevant to the discourse. Contrast is a fundamental principle in sensing and under- standing and leads to the sense of distinctiveness (Twigger-Ross and Uzzell, 1996) and wholeness. A local landmark stands out in an envi- ronment due to its salience in contrast to other places nearby (Raubal and Winter, 2002). In some situations the degree of contrast is low and difficult to be perceived by a human, and as a result a place may be- come vague (e.g., it is often difficult to determine the foot of a moun- tain, or the boundary of a city). The process that a place emerges in cognition is also regarded as similar to how human perceiving a figure (a Gestalt-psychological term meaning a perceptional identifiable ob-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 20 literature review
ject) in comparison to the ground (Levinson, 1996; Talmy, 1983). When referring to a place in different contexts, the contrast set of this place may be different, which may result in different footprints of the place. Such contrast sets of places can either be explicitly mentioned in the discourse of place descriptions, implied, or pre-exist as shared knowl- edge between the descriptor and the recipient. Based on the contrast set theory of places, Vasardani, Stirling, and Winter (2017) provided a conceptual model to interpret the region implied by preposition at. The results derived are context-sensitive, as when places from a contrast set change due to the change of context, the generated re- gion will be different. However, what features determine contrast sets of places in conversations remains unknown, and therefore deriving contrast sets from NL remains a significant challenge.
2.3 spatial language and place description
The relation between space, place and language has also recently at- tracted research attention, addressing issues related to language, ge- ography, and spatial humanities. The driven forces are the increas- ing volume of unstructured textual documents being published on- line (Melo and Martins, 2017), as well as the growing need for place- related information in everyday life. Place information embedded in NL text is relevant to several research areas, such as crowd-sourced geographic information, GIR, and location-based services. Such em- bedded information, once captured, becomes useful as users might greatly benefit themselves by having richer and more locally valid data on place in many situations (Davies et al., 2008). The rapid de- velopment of text mining and NLP techniques makes it feasible to extract information from textual documents, through information ex- traction (IE) techniques (Cowie and Lehnert, 1996) such as named entity recognition (NER) (Nadeau and Sekine, 2007), entity relation extraction (Culotta and Sorensen, 2004), spatial relationship extrac- tion (Khan, Vasardani, and Winter, 2013; Kordjamshidi, Van Otterlo, and Moens, 2011; Liu, Vasardani, and Baldwin, 2014), and event ex- traction (Kim et al., 2009).
2.3.1 Modelling Spatial Language from Text
Several frameworks for modelling spatial-related languages have been proposed in the literature. The GeneralizedUpperModel (GUM) (Bateman et al., 2010) is a comprehensive linguistic ontology for de- coding the semantics of a large proportion of linguistic expressions involving space, not necessarily about places. It is based on the de- scription logic ALCHQ(D) and thus, applies description logic rea- soners for reasoning support. SpatialML (Mani et al., 2010), as a markup language schema, has the goal of automatic annotation of places as well as spatial relationships referred in texts. The spatial role labelling (SpRL) approach (Kordjamshidi, Van Otterlo, and Moens, 2011) focuses on extracting spatial objects (as a superset of places) and
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 2.3 spatial language and place description 21 spatial relationships from text automatically. ISO-Space (Pustejovsky, Moszkowicz, and Verhagen, 2011) is another text annotation language that borrows annotations of static spatial information from SpatialML, along with additional elements such as event, motion, and path. The goal is to create a new standard for spatial-temporal language anno- tation. The major differences between these frameworks in the literature and the place graph (PG) model proposed in this thesis is that, the goal in this thesis is to build a knowledge base from the extracted information from place descriptions. A graph-based model opens op- portunities for new applications that rely on graph traversal, e.g., rea- soning regarding PG as a constrained relational network (Ligozat and Renz, 2004), or knowledge querying based on path connections. Some potential of graph-based representations of spatial information (i.e., SPGs) have already been demonstrated in previous work, such as for creating sketch maps or identifying local landmarks (Belouaer, Bros- set, and Claramunt, 2016; Kim, Vasardani, and Winter, 2016, 2017a). The annotation frameworks introduced above, however, do not di- rectly allow such tasks. This is because it is not trivial to translate these existing annotation frameworks into graphs, as additional chal- lenges including schema design and knowledge mapping need to be addressed. Graph-based databases also have computational advan- tages in storing and querying relational data (Güting, 1994), in com- parison to complex joining operations in relational databases. More discussions about SPG and graph database will be provided in Sec- tion 2.6.
2.3.2 Natural Language Place Description
NL place descriptions occur in everyday communication as a way of encoding and transmitting spatial and semantic knowledge about places between individuals, either to convey or to query spatial in- formation (Vasardani et al., 2013). The information conveyed by such place description is useful for mental sketching of spatial environ- ments, and can be used to, for example, provide navigational instruc- tions or inform the location of events. Place descriptions can be in either verbal or written forms. The web provides a plethora of place descriptions in written form, such as news articles, social media texts, trip guides, and tourism articles. The development of web-harvesting and crowd-sourcing techniques has made the generation of large place description datasets feasible (Kim, Vasardani, and Winter, 2015; Vasardani, Winter, and Richter, 2013). Place descriptions typically pro- vide a qualitative reference system for describing geographic loca- tions, and consist essentially of references to places and their qualita- tive spatial relationships, e.g., “The courtyard is on the campus, beside the clock tower”, describing the location of the courtyard in relation to two other places. Semantic information about places may also be conveyed through place descriptions through describing place types, events, and human activities that happened at places.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 22 literature review
The structure of place descriptions has been studied in linguistics (Jarvella and Klein, 1982; Schegloff, 1972) and spatial cognition com- munities (Couclelis et al., 1987; Jarvella and Klein, 1982), and recently in the domain of GIScience as well. Place descriptions allow success- ful communication of spatial knowledge. However, a computer has difficulties understanding such descriptions (Winter et al., 2010), and currently there is little research attention in modelling and utilizing place knowledge from place descriptions. Place descriptions are ex- pressions that refer to places either by their names (‘Southern Cross Station’), by their types or functions (‘the train station’) (Bennett and Agarwal, 2007; Vasardani et al., 2013), or possibly in other vernacular forms (‘the place we met yesterday’). Spatial relationships are often expressed as prepositions (e.g., on, in), but can also be a verb (e.g., sur- rounding), a phrase, or even be implicit (Vasardani et al., 2013). Many place descriptions apply a hierarchical structure in terms of spatial granularity or salient (Couclelis et al., 1987; Richter et al., 2013) which reflects hierarchical structure in cognitive spatial representations. Place descriptions typically contain locative expressions (Her- skovits, 1985) that describe places in relation to other places with their locations better known, e.g., ‘the building is behind the State Library’ describes the location of the place building in relation to a landmark State Library. A locative expression consists of a locatum, the reference to a place that is to be located, a relatum, the reference to a place that is already located, and a spatial relationship between the two. Locative expressions can help locating places referred by references that are not official, unambiguous place names. The locatum and re- latum in a locative expression have also been called as a locatum and a reference object in the spatial semantic parsers’ literature (Coyne, Sproat, and Hirschberg, 2010; Kordjamshidi, Van Otterlo, and Moens, 2011; Kordjamshidi et al., 2011). In other disciplines such as cognitive linguistics and psychology, a spatial entity whose location is of rele- vance has also been called a trajector (Lakoff, 1987; Langacker, 1987), a figure (Levinson, 1996; Talmy, 1983), or a general term referent. It has also been called a target as the subject of a spatial relation, and the corresponding object of the relation is called a landmark (Vandeloise, 1991). The term landmark has also been used to describe the refer- ence entity in relation to which the location of a trajector is specified (Langacker, 1987). Generally, a relatum, or a landmark, is an entity whose location is better known thus in relation to which the location of the locatum can be specified. A locatum defined by these terms are generally similar: it may be static or dynamic; a person or an object whose location is of relevance. In the literature there are also minor differences in the definition of these two terms. For example, an event can become a locatum as specified in (Langacker, 1987), and a deic- tic centre as a point of reference is also regarded as a landmark in (Langacker, 1987). NL descriptions about places have been regarded as one of the emerging modes of interacting with GISs and location-based services. It also provides a unique opportunity for the generation of a rich data source about places from human spatial knowledge. It has been
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 2.4 qualitative spatial relations 23 postulated that such information might be leveraged to create a place- based information system for modelling and utilizing human knowl- edge about place (Elwood, Goodchild, and Sui, 2013). In order to set up such systems, three major research problems can be identi- fied: IE through NLP, information modelling, and knowledge utilisa- tion. For the first task, techniques have been developed, with exam- ples introduced above. For this thesis, off-the-shelf NLP techniques will be used for IE purpose. This thesis focuses on the second and third tasks, i.e., modelling and utilizing knowledge extracted from place descriptions. Methods have also been developed for harvesting place description from websites with the help of application program- ming interfaces (APIs) and web crawlers (Kim, Vasardani, and Winter, 2015), which allow the convenient generation of large place descrip- tion datasets.
2.4 qualitative spatial relations
Qualitative spatial relations, such as near, in front of, and north of, are relevant to this thesis from two aspects. For the reasoning phase, they are used for relational inference as well as consistency maintenance. For the georeferencing phase, the approximate locations of places are inferred considering the relationships to other places as anchors. Ac- cordingly, this section surveys research into how qualitative spatial relations are modelled in the literature, in particular for qualitative spatial reasoning (QSR) calculus as well as for quantification purposes.
2.4.1 Qualitative Spatial Representation and Reasoning
People use qualitative spatial relations often when describing places and spatial objects. When the descriptions came from memory, they are based on the cognitive spatial representations, or mental models, of the environment (Kuipers, Tecuci, and Stankiewicz, 2003; Tversky, 1993). People also rely on mental models for reasoning tasks in simple scenarios, although such models often lead to incomplete and biased results (Knauff, Rauh, and Schlieder, 1995). In English, such qualita- tive spatial relations are often expressed by spatial prepositions, and the semantics of such prepositions, in terms of their spatial, temporal, and geometrical meanings, have been studied in linguistics (Gärden- fors, 2014; Landau, 1994; Tenbrink, 2007; Zwarts and Winter, 2000). Spatial prepositions can be identified and extracted from texts using a parser. In the Artificial Intelligence community, qualitative spatial rela- tions have been studied for QSR (Cohn and Hazarika, 2001; Freksa, 1991; Ligozat and Renz, 2004; Retz-Schmidt, 1988; Zimmermann and Freksa, 1996), where they are formalized in logical or algebraic calculi. QSR is an interdisciplinary, broad research area with shared interests from several communities, such as cognitive science, spatial cogni- tion, and robotic navigation. So far formal reasoning frameworks are provided mainly within four families of relationships: topological re-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 24 literature review
lations, cardinal direction relations, relative direction relations (for which the reference frame is relevant, compared to cardinal direction relations), and qualitative distance relations. For the topological relation family, two competing models have been developed, the region connection calculus based on first-order logic (Randell, Cui, and Cohn, 1992), and the intersection model based on point set topology (Egenhofer and Franzosa, 1991). Both lead to the same set of eight topological relationships between two simple regions: disjoint, touch, overlap, equal, covers, covered_by, contains, and inside. Experiments in order to calibrate the meaning of spatial predicates (in English) show strong correlation between human cat- egorizations and these formally distinguished relations (Mark and Egenhofer, 1994; Mark et al., 1995). For the cardinal direction relation family, Frank (1992) distin- guished three models for a cardinal direction calculus: a cone model and a half-plane model for point-like objects, and a neutral zone model for spatially extended objects. Other models also exist, such as the internal cardinal direction model (Liu et al., 2005). Models for car- dinal direction relations also come in different granularities, includ- ing models with four relations: north, south, east, and west, and with eight relations (the four relations together with northeast, northwest, southeast, and southwest). Calculi based on cardinal direction relations have also been developed (Freksa, 1992; Ligozat, 1998). For the relative direction relation family, a typical four-valued sys- tem of in_front_of, behind, left and right are distinguished considering the reference frame, which can be linked to the speaker (deictic), to a geographic entity (intrinsic), or to situational context (extrinsic) (Retz- Schmidt, 1988). The reference frame in NL expressions, however, is typically hidden in context and not easily capturable. Similar to car- dinal direction relations, relative direction relations also come in dif- ferent granularities (Moratz, Dylla, and Frommberger, 2005). Popular formal models of relative direction relations include the dipole model (Schlieder, 1995), the double-cross calculus (Freksa, 1992), and more (e.g., see Wolter and Lee, 2010; Wolter, 2009). More recently, a formal model of relative direction under different reference frames is pro- vided (Tenbrink, 2011; Tenbrink and Kuhn, 2011). A framework for relative direction reasoning under different reference frames has also been proposed (Hua, Renz, and Ge, 2018). A review of current formal representations is given as well (Freksa, Ven, and Wolter, 2018). Lastly, the qualitative distance relation family, with relations such as near and far, has also been formalized (Duckham and Worboys, 2001; Frank, 1992). However, the family is relatively under-researched due to the fact that their meaning is heavily context-dependent and challenging to be captured (Duckham and Worboys, 2001). Frank (1992) discussed two-step (close, far), three-step (close, middle, far) and multi-step systems. Those systems consider qualitative distance as mappings into intervals that form a partition of the positive real numbers. In place descriptions, however, the defined reasoning rules of the model may not apply due to the flexibility as well as context- dependency of NL.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 2.4 qualitative spatial relations 25
Calculi have been defined associated with these proposed models for QSR tasks, such as robotic navigation and consistency maintenance. Consistency maintenance is the interest of this thesis. A composition operation (denoted by ⊕) is defined as the set of possible relations SA,C between features A and C, given the relations rA,B and rB,C. SA,C can be a single relation rA,C, a set of relations, or the universal relation any. Only an inconsistent composition can lead to an empty set. Each of the families above has been investigated for composi- tion, and composition tables of topological (Düntsch, 2005; Egenhofer, 1994; Renz, 2002), cardinal (Frank, 1996), relative (intrinsic) directions (Schlieder, 1995), or qualitative distance relations (Frank, 1992; Hong, Egenhofer, and Frank, 1995) have been developed. Some calculi have also been implemented as toolboxes (Wolter and Wallgrün, 2012). A few reasoning models with calculi combining multiple families have also been proposed (Cohn et al., 2014; Gerevini and Renz, 1998, 2002). These proposed frameworks in the field of QSR, including spatial relation models and associated calculi, are typically for objects with simple, crisp geometries and may not be appropriate for places. It remains to be seen whether and how these frameworks apply for spatial reasoning of human spatial knowledge about places as well.
2.4.2 Modelling Qualitative Spatial Relations Quantitatively
Existing webmapping applications typically treat different types of spatial relationships in queries equivalent to near (Winter and Tru- elove, 2013). The semantics of these spatial relationships, such as in front of or at, are ignored, since they are currently difficult for com- puters to interpret. Although such simplification often leads to useful enough results, the process of interpreting the query is essentially different from the way a human would proceed. A recipient of a de- scription must be able to distinguish the semantics of different spatial relations in order to correctly decode the description. For example, in the description “I am in front of the building that is behind the State Library”, the location of the descriptor cannot be correctly located without considering the semantics of two spatial relationships men- tioned. Spatial relations are important both for place reference disam- biguation, as well as spatially anchoring places by limiting their likely locations. Together, spatial relations and place referents provide suf- ficient information for conveying locative information about places. Thus, even though some place references are ambiguous or vernacu- lar in descriptions, recipients can usually locate them. However, due to the vagueness and flexibility of the uses of spatial relations in NL (Fogliaroni and Hobel, 2015; Liu et al., 2009b; Schw- ering and Raubal, 2005), modelling spatial relationships for compu- tations in GISs and spatial databases remains a significant challenge. People use locative expressions to describe a vague location in rela- tion to another place with location known. For example, ‘10km east of Berkeley’ refers to some unspecified place at a 10km distance in a particular direction from a known place called Berkeley. Wieczorek,
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 26 literature review
Guo, and Hijmans (2004) developed a point-radius method using car- dinal directions and (imprecise) metric distance relationships for such expressions. Associated uncertainties such as coordinate-, distance-, and direction-imprecision are calculated to derive a probabilistic un- certainty field representation. The methodology was later modified by applying a probabilistic distribution model (Guo, Liu, and Wiec- zorek, 2008). Liu et al. (2009b) go a step further by adding topological and qualitative distance relationships in the model, yet the param- eters require manual configuration by the user when applying the models. In addition, these probabilistic models assume that the loca- tions of the relata in locative expressions are known. In comparison, this assumption is not adopted in this thesis. Fu, Jones, and Abdel- moty (2005b) assign different search radii for relations such as near or north based on the size of the places for spatial querying, and the distance parameters are again empirically adjusted. Other studies attempt to quantify qualitative spatial relations us- ing data-driven methods. Delboni et al. (2007) focus on determining semantic equivalence of distance relations for query expansion pur- poses. Hall, Smart, and Jones (2011) quantify spatial relations in terms of distance and orientation. Skoumas et al. (2016) derive probabilis- tic models for spatial relations and choose only major metropolises as their case study. Derungs and Purves (2016) use web n-grams to model vague spatial relation concepts. These previous studies typi- cally focus on spatial relations from certain contexts, such as spatial relations linked to major cities. In comparison, this thesis aims at gen- eralizing the models to make them scalable to different places for georeferencing purposes, as well as to contextualize these models to increase their flexibility to accommodate different contexts. There are studies suggesting to bring in context for interpreting qualitative spatial relations. However, these models either remain at conceptual level, or require knowledge that is not available in this thesis. Cai proposed a framework to contextually model geospatial data considering tasks and transportation tools (Cai, 2007). However, such knowledge is not available in this thesis for place descriptions. Wallgrün, Klippel, and Karimzadeh (2015) propose identifying key context features affecting human usage of spatial relation expres- sions, in order to produce contextualized models for answering spa- tial queries. Despite the fact that no implementation is yet provided, the goal is similar to this thesis. Yao and Thill studied context explic- itly by investigating how contextual features could affect the inter- pretation of proximity measures such as near and not so far (Yao and Thill, 2005). However, most of such contextual features, e.g., famil- iarity with the area, financial and time budget, network connectivity, and personal characteristics, may not be contained by or easily ex- tracted from place descriptions. Instead, contextual features that are derivable from place descriptions are preferable in this thesis.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 2.5 georeferencing place from textual documents 27
2.5 georeferencing place from textual documents
Geo-referencing is the process of associating places with geographic locations, and the process is often with the help of some external source of knowledge, typically a gazetteer. The task for locating places referred in textual documents is typically referred as toponym resolution (TR) (Leidner, 2007), and extensive methods have already been proposed. This section is divided into two parts. The first part reviews the general approaches for TR from different categories, and a discussion of their applicability and suitability for the task of this thesis is given. The second part particularly focuses on approaches proposed for georeferencing fine-grained places and non-gazetteered place references.
2.5.1 Toponym Resolution
TR is one of the core tasks for building GIR engines as well as de- termining the geo-focus of textual documents (Anastácio, Martins, and Calado, 2009; Laere et al., 2014; Purves et al., 2007; Wing and Baldridge, 2011). Geotagging systems based on TR have been devel- oped as well for determining the geographical location embedding of textual documents (Adelfio and Samet, 2013a; Amitay et al., 2004; Lieberman, Samet, and Sankaranarayanan, 2007; Teitler et al., 2008; Woodru and Plaunt, 1994). The goal of TR is to both recognize place names from textual doc- uments (e.g., web texts or historical documents) and link them to ge- ographic locations. Therefore, TR essentially comprises two subtasks, namely toponym recognition, and disambiguation. Alternative names for the two subtasks, e.g., geoparsing and geocoding, have also been used interchangeably as well. Toponym recognition is typically done by NLP techniques based on NER and gazetteer matching (e.g., Lei- dner and Lieberman, 2011; Loureiro, Anastácio, and Martins, 2011), and non-gazetteered place references are ignored. Moreover, some names are semantically ambiguous (e.g., the word Melbourne may be either a place name or a person name and thus has to be semantically disambiguated) and have to be resolved as well. The second subtask, toponym disambiguation, is to map each recognized place name to its actual, unambiguous geographic location. This step is required as toponyms are rarely unique and may have more than one corre- sponding gazetteer entries, e.g., GeoNames lists 14 populated places named Melbourne world-wide. For the disambiguation task, extensive approaches have been proposed that can be classified into several cat- egories, and they are often used in conjunction with various heuris- tics. Since this thesis relies on a previously-developed parser for ex- tracting place references from descriptions, toponym recognition is outside the scope and only toponym disambiguation is considered. Toponym disambiguation approaches are typically done by search- ing for context places, i.e., other place names occurred in the same document, and computing the likelihood of each of the candidate
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 28 literature review
gazetteer entry corresponding to the place to be disambiguated. The likelihood is computed as a score based on heuristics (or a combi- nation of heuristics (Derungs and Purves, 2014; Leidner and Others, 2004)) using knowledge such as locations or containment relation- ships. For example, if two place names Melbourne and Florida occur to- gether in a document, then the place name Melbourne is more likely to be corresponding to the gazetteer entry ‘Melbourne, Florida, United States’ rather than ‘Melbourne, Victoria, Australia’, since in the later case the disambiguated locations of Melbourne and Florida are closer to each other. A comprehensive list of heuristics is summarized by Lei- dner (Leidner, 2007) as used by previous TR studies, such as choosing the one with largest population, or being a singleton capital. Depending on the knowledge leveraged, disambiguation methods can generally be classified into map-, knowledge-, as well as machine learning-based (Buscaldi, 2011). Map-based approaches rely mainly on the locations of the gazetteer entries of places names from a docu- ment, and use heuristics such as minimum point-wise distance, min- imum convex hull, or closest to the centroid of all entries locations for disambiguation (e.g., Amitay et al., 2004; Smith and Crane, 2001; Zhang et al., 2012). Previous studies that focus on disambiguating fine-grained places (e.g., Derungs, Palacio, and Purves, 2012; Moncla et al., 2014; Palacio, Derungs, and Purves, 2015), are largely based on map-based approaches as well. Knowledge-based methods lever- age external knowledge of places such as containment relationships, population, or prominence (e.g., Adelfio and Samet, 2013b; Buscaldi and Rosso, 2008b; Karimzadeh et al., 2013). Machine learning-based approaches rely on creating language models from training data that represent the likelihood of seeing each of the places from the same document associated with a place to be disambiguated (e.g., Garbin and Mani, 2005; Roberts, Bejan, and Harabagiu, 2010; Smith and Mann, 2003). There are also more recent approaches that not only consider other place names from the same documents, but also other words as well (Cheng, Caverlee, and Lee, 2010; Liu and Inkpen, 2015; Roller et al., 2012; Wing and Baldridge, 2014). These approaches have the advantage of using non-geographical context words such as events, person names, or organization names to assist disambigua- tion. Hybrid and bootstrapping approaches combining different mod- els also exist (Adams and Janowicz, 2012; Cheng, Caverlee, and Lee, 2013; Han, Cook, and Baldwin, 2014). Approaches for TR typically focus only on places that are gazetteered (i.e., officially indexed in gazetteers) and of spatial gran- ularities that are larger or equal to suburb- and city-level, and ignore references to places that are not. Not as much attention and effort has been spent on developing approaches for resolving other types of places or references to places. The reason is that prominent and gazetteered places are usually more important for their applications. In comparison, this thesis additionally considers several new chal- lenges, which will be discussed in the next section.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 2.5 georeferencing place from textual documents 29
2.5.2 Georeferencing Fine-Grained Places and Non-Gazetteered Place Ref- erences
The selection of the disambiguation approach is usually task- and data-dependent (Buscaldi, 2011). For the task of georeferencing all ex- tracted references to places from everyday place descriptions in this thesis, existing approaches from TR are typically not applicable due to several reasons. Place descriptions often contain place names of fine-grained features, which are often significantly more frequent and more similar to each other. Place descriptions also often contain place referents that cannot be found in a gazetteer, such as synonyms or place type references, and some of these places may not have corre- sponding entries in gazetteers, such as references to places that are from too fine-grained environments to be captured by gazetteers. Some studies aim to overcome these limitations and further geo- reference places that are fine-grained, referred by non-gazetteered place references, or not captured by gazetteers. Palacio, Derungs, and Purves (2015) developed an approach for disambiguating fine- grained toponyms based on Euclidean distance and topographic sim- ilarity to anchor toponyms (ones that are easily resolvable) from the discourse. However, the approach requires the toponyms to be dis- ambiguated being both gazetteered and associated with type infor- mation. Geiß et al. (2015) proposed a graph-based approach based on constructing a weighted network for co-occurring toponyms from Wikipedia. Later, Spitz, Geiß, and Gertz (2016) leverage the network for toponym disambiguation based on computing a ranking of can- didates by their co-occurrence with other toponyms mentioned in the documents. However, the network must be georeferenced before- hand, which means it cannot be used to resolve new toponyms that did not appear in the network. Finally, some studies focus on develop- ing gazetteer-independent approaches. Language models have been used for georeferencing toponyms and documents (Cheng, Caverlee, and Lee, 2010; Roller et al., 2012; Wing and Baldridge, 2014). These methods typically discretize the earth into cells, and train discrimi- native classification models to associate documents with these cells. Then, similarity scores are computed to decide which cells are best corresponding to a given test documents. Delozier, Baldridge, and London (2015) developed a gazetteer-independent approach that cal- culates the likelihood of seeing a word at a certain location, and find points of strongest overlap for a toponym and context words. How- ever, the mean and median distance errors of these language model- and grid-based methods are generally at the scale from hundreds to thousands of kilometers, which make these approaches less useful for locating or disambiguating fine-grained places. Compared to knowledge- and machine learning-based disambigua- tion approaches, map-based approaches seem relatively robust for the disambiguation of fine-grained places as they only require knowledge of the locations of ambiguous candidate entries. Map-based disam- biguation approaches are typically based on clustering and share the same assumption that most places in a text are connected and likely
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 30 literature review
nearby. The input to these clustering algorithms are the locations of all ambiguous candidate gazetteer entries of all place names from a document, in the form of a point cloud. These places to be disam- biguated are expected to be within clusters that are geographically minimal. Several approaches have been developed. The overall mini- mum distance heuristic (Amitay et al., 2004; Habib and Keulen, 2012; Leidner, Sinclair, and Webber, 2003) aims at selecting gazetteer en- tries so that they are as geographically close to each other as possible. The closeness is measured by, for example, the average location-wise distance or area of the convex hull of these locations. The centroid based heuristic (Buscaldi and Rosso, 2008a; Smith and Crane, 2001) computes the geographic focus (centroid) of all ambiguous entry loca- tions, and calculates the distance of each entry location to it. A thresh- old is defined to exclude entry locations that are too far away from the centroid, probably in an iterative manner. Finally, for each place name, select the entry that is closest to the centroid for disambigua- tion. The minimum distance to unambiguous referents heuristic (Buscaldi and Magnini, 2010; Smith and Crane, 2001) identifies unambiguous place names, i.e., place names with only one gazetteer entry, or place names that can be easily disambiguated based on some heuristics, and then use a scoring function for the disambiguation of the remain- ing ambiguous entries, such as based on average minimum distance to those unambiguous entry locations, or weighed average distance considering times of occurrence in document or textual distance. Compared to the previous clustering methods defined in ad-hoc manners, clustering algorithms from the data-mining community for dividing data into meaningful groups of objects have been used as well. The DBSCAN algorithm (Density Based Spatial Clustering of Applications with Noise) is a density-based method that relies on two parameters: the neighborhood distance threshold ε, and the min- imum number of points to form a cluster MinPts. Moncla et al. (2014) use DBSCAN for the disambiguation of fine-grained toponyms from hiking descriptions. Using DBSCAN requires a-priori knowledge of the input data to determine the parameters, and the parameters in their case were empirically adjusted. Place descriptions, however, have potentially various conversational contexts, and thus vary in spa- tial coverage as well as distances between the places being mentioned. DBSCAN is not suitable for disambiguating places from place de- scriptions, as there is no easy way to determine the parameter values without priori knowledge of each input description. There are also many other clustering algorithms defined in the data-mining commu- nity for statistics, pattern recognition, and machine learning purposes. However, the performances of these algorithms for the task of place name disambiguation have not been tested before. Some algorithms are less parameter-sensitive and thus may not require pre-knowledge of the input data. A review of these clustering algorithms is given by (Berkhin, 2006), and some of them will be introduced and compared later in Chapter 5. Finally, despite the extensive studies on modelling qualitative spa- tial relations, as well as the expected efficacy of these models being
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 2.6 graph representations of spatial knowledge 31 leveraged in georeferencing place from text, spatial relation models are hardly considered in TR tasks. In a recent review of current ap- proaches for geocoding textual documents, spatial relationships other than hierarchical containment are not discussed (Melo and Martins, 2017). Although Moncla et al. (2014) discuss the possibility of lever- aging spatial relationships from NL to approximately locating non- gazetteered toponyms, their actual implementation is still limited to using the convex hull and circumscribed circle of anchor places for approximate localization.
2.6 graph representations of spatial knowledge
A graph G = (V, E) is a representation of vertices (or nodes) con- nected by edges (West and Others, 2001). A graph may either be undirected, meaning an edge connecting two nodes has no direction, or directed, and it can also have properties associated with nodes and edges. Graphs have been used in various disciplines, such as computer science, linguistics, biology, and social science, for mod- elling relationally-connected data. With the development of seman- tic web and linked data (Berners-Lee, Hendler, and Lassila, 2001; Bizer, Heath, and Berners-Lee, 2011; Bollacker et al., 2008), graph- based representations of knowledge and graph databases have been used by researchers as well as industrial companies such as Facebook and Google. For instance, social network graphs have been used for analyzing connections between individuals and groups (Wasserman and Faust, 1994), and knowledge graphs have been used for tasks such as information retrieval or question answering (e.g., Dalton, Di- etz, and Allan, 2014; Yao and Van Durme, 2014). In the domain of GIScience, graphs have been used for representing various types of spa- tial knowledge, such as for modelling transportation networks (e.g., streets networks and airline networks) for route planning and analy- sis purposes, or for modelling spatial cognition for robotic wayfind- ing (Kuipers and Byun, 1991). Vasardani et al. (2013) regard place descriptions as place references and spatial relationships embedded in locative expressions, which can be extracted using a parser and modeled by triplets. A triplet is the representation of a locative expression by locatum (L), relatum (R) and their relation (r). The representation is defined similarly as in SpRL (Kordjamshidi, Van Otterlo, and Moens, 2011). For example, the description “The courtyard is on the campus, beside the clocktower" can be modelled in the form of triplets:
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 32 literature review
exclude false positives such as metaphorical, temporal, and fictive mo- tion ones. Although these parsers typically incorporate mechanisms for this purpose, it is noted in some of these studies that such locative expressions are still one of the major sources of false positives. Other sources include informal locative expressions and complex locative expressions that are difficult to be decode and extracted by the devel- oped parsers. A SPG can then be constructed based on the extracted list of triplets (Vasardani et al., 2013). Each triplet is stored as two nodes, one each for the locatum and the relatum, and an edge in between for the spa- tial relationship. Spatial relationships from NL expressions are usually qualitative, and standardized in their vocabulary for their representa- tion in a SPG, e.g., ‘northern’ is normalized by an edge north. The edges in the graph are directed from locatum to relatum due to the asymmetry of spatial relationships. Such a SPG constructed from loca- tive expressions is also called a PG. The two example triplets from above can be used to create a simple PG, as shown in Figure 3. PGs have been leveraged to create sketch maps (Kim, Vasardani, and Win- ter, 2016) and to identify landmarks (Kim, Vasardani, and Winter, 2017a). Compared to the object- (e.g., place in gazetteer) and field- based models for places (e.g., density surfaces for vague places (Gao et al., 2017a; Hollenstein and Purves, 2010; Jones et al., 2008a)), a PG additionally captures the network dimension (Kuhn, 2012) of places by their co-occurrence and spatial relationships in descriptions.
Figure 3: A simple place graph representing the spatial references “the courtyard is on the campus" and “the courtyard is beside the clock- tower".
When a PG is constructed from collective place descriptions, one challenge is to identify nodes referring to identical places. The refer- ences to the same place may be made in different references, and thus, nodes referring to identical places should be identified and merged. For example, if the two triplets are from two different dis- courses, the graph in Figure 3 cannot be created unless the two court- yard references are detected to be referring to the same place. Kim, Vasardani, and Winter (2017b) developed a comprehensive approach for graph merging by identifying place identity considering reference string similarity, reference semantic similarity, and similarity of spa- tial relationships to other nodes (places). Sufficiently similar places are stored as a single node, i.e., each node has a unique identifier and potentially multiple place references. Note that this task is differ- ent from toponym disambiguation, as node merging does not require the place references to be gazetteered, and the process will not link places to locations either. Therefore, there can be multiple edges rep- resenting different spatial relationships between the two place nodes, because the same places may be referred to multiple times, or in sev- eral place descriptions. For each type of relationship, at most one in-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 2.6 graph representations of spatial knowledge 33 stance is stored between any pair of nodes as an edge, and additional instances will be regarded as duplicates and discarded. Other than PGs for place descriptions, SPG representations have also been developed for modelling route descriptions. The major differ- ence is that route descriptions typically involve human actions (e.g., movements indicated by motion verbs), while place descriptions tend to focus on describing spatial configurations of places through their mutual spatial relationships. Although place descriptions and route descriptions are often mixed in usage. Route graph models for out- door environment have been proposed that are typically constructed from triplets as well, similar to a PG, e.g.,
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 34 literature review
2.7 summary
This chapter describes relevant work that are related to the identified four major challenges of this thesis, i.e., place knowledge modelling, QSR, place georeferencing, and place knowledge querying. It has been argued that existing GISs and spatial databases are ma- ture in representing space, but limited in representing place. Some existing place models have been introduced in Section 2.2. Place de- scriptions provide a rich source of human knowledge about places and seems promising to be used for creating a place knowledge base. Although approaches have been proposed for decoding spatial lan- guage from text, as discussed in Section 2.3, constructing a place knowledge base using such decoded knowledge is not considered by these studies. The task requires additional challenges to be resolved such as knowledge modelling as well as identifying the utility of the knowledge base. Spatial relations are commonly used in place descriptions for de- scribing the relative locations of places, which are based on the cog- nitive spatial representations of the environment. The semantics of spatial prepositions from NL has been studied in linguistics, and spa- tial relations are modelled in the field of QSR by families with asso- ciated calculi for reasoning tasks, as introduced in Section 2.4. As a knowledge base, the capacity of reasoning with the stored spatial relationship knowledge is useful for consistency maintenance. How- ever, spatial relationships in place descriptions are often from differ- ent families, and the variety and flexible use of NL spatial relations may not apply these formal models. The task of georeferencing places from textual documents is typ- ically solved by the proposed approaches in the community of TR, which have been discussed in Section 2.5. The selection of the disam- biguation approach is usually task- and data-dependent. Since every- day place descriptions often contain place references of fine-grained features, or references that cannot be found in gazetteers, most exist- ing TR approaches are not suitable or directly applicable for the task in this thesis: georeferencing all extracted references to places from ev- eryday place descriptions. In addition, although spatial relationships are expected to be useful for georeferencing places from text, they are hardly considered in previous approaches. Finally, the concepts of knowledge graphs, SPGs, and graph databases have been explained in Section 2.6.A SPG database seems naturally suitable for modelling places, place properties, as well as spatial relationships extracted from place descriptions. Such a database can be regarded as a place knowledge base constructed from collective place descriptions. The database is selective, and does not aim for completeness as existing GISs and spatial databases typical do. In order to utilize the captured knowledge in application scenarios, the database should ideally support querying tasks. Thus, the last challenge of this thesis is to study what types of queries are answer- able by such a database, as well as how they can be queried.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] PLACEKNOWLEDGEMODELLING 3
Place descriptions provide a rich source of spatial and semantic knowledge of places. Such knowledge may be complementary to the knowledge that traditional geographic information systems (GISs) and authoritative databases provide, and thus, it can be used to construct a place knowledge base and facilitate a range of place-related applica- tions. Although several annotation models have been proposed previ- ously for decoding spatial language, as introduced in Chapter 2, con- structing a collective place knowledge base from place descriptions is a problem from a different perspective and requires a different model. It is also not a trivial problem to translate these language models to a knowledge base model, since issues including schema design and knowledge mapping need to be addressed. The purpose of this thesis is to construct such a knowledge base and use it for reasoning, georef- erencing, and querying tasks, while these proposed language models do not yet allow such tasks. This chapter proposes a place graph database (PGD) model for mod- elling the spatial, non-spatial, and contextual knowledge extracted from place descriptions. The model restructures and largely extends a prior place graph (PG) model (Vasardani et al., 2013) (referred as the basic PG in the remaining part of the thesis) and overcomes a num- ber of limitations. The model is implemented using a graph database with description datasets, and the implemented database will be used in the later chapters for experiments. This chapter is based on content from a published paper (Chen et al., 2018). The contributions of the corresponding (first) author in- clude problem conceptualization (through discussions with other au- thors), literature review, identifying knowledge to be modelled, the design and implementation of the PGD model, experiments, and pa- per writing. Experiments from the paper have been reorganized and will be presented in Chapters 4, 5, and 6 respectively.
3.1 introduction
Place descriptions occur in verbal and textual communication as a way of encoding and transmitting spatial and semantic knowledge about places between individuals. Information conveyed by a place description, such as place references, spatial relations, and human activities, represents the locative and semantic knowledge of the de- scriptor about the referred places. It has been postulated that such information might be leveraged to create a place-based information system for modelling and utilizing human knowledge about place (Elwood, Goodchild, and Sui, 2013). Place descriptions have been regarded as a qualitative reference system for describing geographic locations, and consist essentially of
35
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 36 placeknowledgemodelling
references to places and their qualitative spatial relationships. Such information extracted from place descriptions has been used to con- struct a (basic) PG (Vasardani et al., 2013). Figure 4 shows an example basic PG, which consists of places as nodes and spatial relationships as labeled edges. The input to created such a basic PG are triplets of locative expressions extracted from place descriptions.
Figure 4: An example basic place graph with node size corresponding to node degree, and edge size corresponding to number of relation- ships between the linked nodes. Multiple relationships between two nodes are represented by only one edge.
However, the triplets used to build such a basic PG are stripped off of much of their conversational contexts. It is, therefore, possible to find incompatible information, especially if the graph is constructed from combining place descriptions with different conversational con- texts. For example, it is perfectly possible to collect seemingly contra- dicting triplets such as
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 3.2 extending the place graph model 37 graph traversing, without filtering for context, inferences, or seman- tics, while the interpretation of spatial relationships such as relative directions or qualitative distances is limited. This chapter reorganizes, revises and extends the basic model, in order to capture information from place descriptions that is useful, but lost during the modelling process. The remainder of the chapter is structured as follows. In Section 3.2, the chapter commences with identifying several types of information that are not captured in the basic PG model (Section 3.2.1). Then, the extended model is presented by introducing different types of nodes, edges, as well as properties associated with them (Section 3.2.2). The model is implemented as a graph database that allows efficient query- ing for experiments in the later phases of this thesis (i.e., reasoning, georeferencing, and querying). The implementation details, including a description of the datasets used, the processing procedure, as well as a comprehensive example of graph construction, are given in Sec- tion 3.3. The constructed database will be used in the later chapters for experiments. Section 3.4 summarizes and concludes this chapter.
3.2 extending the place graph model
This section first analyses types of information that are not captured in the basic PG model, as well as the tasks for which they matter. Then, an extended PG model that caters for this information is introduced.
3.2.1 Information not Captured in the Basic Place Graph Model
Figure 5 shows the unified modeling language (UML) diagram of the basic PG model, in which places are treated as nodes connected by spatial relations as edges. The types of information identified in this section below are not considered in the basic PG model. Many of them provide contextual knowledge, which could affect the in- terpretation of other information communicated in place descrip- tions (e.g., spatial relations), and thus should be captured. The def- inition of context is task-specific; in this chapter the categorization proposed by Wolter and Yousaf (2018) of description-, environment-, and human-dependent contexts is considered. For instance, near can refer to different distances according to factors such as spatial granu- larity, or places relevant to the discourse (Winter and Freksa, 2012) (description-dependent context). Certain relations require informa- tion from the environment in order to be interpretable, such as refer- ence frames (environment-dependent context). Individual places and spatial relations can also have different meanings for different people (human-dependent context).
3.2.1.1 Place Semantics and Characteristics Place descriptions sometimes contain non-spatial information about places, such as their types (e.g., ‘the room is a lecture theater’), the activities they afford (e.g., ‘having seminars and lectures’), the things
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 38 placeknowledgemodelling
Figure 5: UML diagram of the basic place graph model.
they equip (e.g., ‘the room has a projector’), as well as their char- acteristics (e.g., ‘old, large’) (Alazzawi, Abdelmoty, and Jones, 2012; Scheider and Janowicz, 2014). Place semantics and affordances have been used for characterizing places and enabling place-based search as well as analysis (Adams and Janowicz, 2015; Alazzawi, Abdelmoty, and Jones, 2012; Janowicz, 2006). Different places may have the same affordances, and one place may have multiple affordances according to individuals or time pe- riods. The way that a gazetteer categorizes places does not always align with the way people regard these places, despite that such cat- egorization is useful in many applications. Capturing semantics and characteristics of places in a PG could provide additional dimensions for tasks such as georeferencing, identical-place matching, and query- ing. In place descriptions, these types of information are often ex- pressed in certain patterns, e.g., as adjectives, nouns followed by words such as ‘is’ and ‘has’, or as verb phrases. Such patterns can be recognized using a trained parser, and the feasibility of creating such a parser has been demonstrated in previous research (Alazzawi, Abdelmoty, and Jones, 2012; Hobel and Fogliaroni, 2016).
3.2.1.2 Places and Relationships from Discourse and Their Sequential Or- der of Appearance Places referred to in different discourses provide contextual knowl- edge for interpreting spatial relations and locating places. For in- stance, near in the description ‘the building is near the Flinders Street Station’ can be interpreted differently in terms of distance, depend- ing on the spatial context (the geographic extent the description is embedded in), e.g., the limited area around the station, or the whole Melbourne CBD. Such a spatial context can be inferred by looking at the places mentioned in the same discourse. Other than places, spatial relationships from the same discourse provide contextual knowledge as well. For example, relative direc- tion relationships can be used to infer the reference direction used by the descriptor, especially when using local landmarks as relata. The inferred reference directions can help with interpreting other relative direction relationships in the discourse, and thus be used to locate places as locata of these relationships. The order of appearance of places and relationships in a place de- scription should also be preserved. For example, descriptors often switch the level of spatial granularity monotonically, e.g., changing from city-level to district-level (Richter et al., 2013). Such changes in context cannot be detected without recording the order of appear-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 3.2 extending the place graph model 39 ance of place references and spatial relationships. Similarly, reference directions can also change within a description, such as at turns, and affect the interpretation of subsequent relations. Storing sequential order also helps linking different place refer- ences that are referring to the same place. Definite references such as ‘the building’, which refers to a building described previously in the discourse, can be ambiguous without sequential appearance infor- mation, if there were multiple buildings mentioned in the discourse. Information about places and relationships from the same dis- course, as well as their sequential order of appearance is not mod- elled in the basic PG model. Triplets from different descriptions are merged without any indexing mechanisms for future separation. The two types of information can be obtained directly without requiring an additional parser; the challenge is how to modify the PG model in order to store this knowledge.
3.2.1.3 Reference Frame and Direction
The basic PG model does not capture spatial reference frame and refer- ence direction information (Levinson, 1996; Tenbrink, 2011). Anchor- ing relative direction relations is, thus, problematic, as it is unknown which directions are being referred to. It is also difficult to perform qualitative spatial reasoning (QSR) or to interpret seemingly contra- dicting direction relations, as in the example
3.2.1.4 Non-Binary Relationships Non-binary spatial relationships, e.g., a is between b, c, or a is sur- rounded by b, c, d, involve more than two places thus cannot be repre- sented by the aforementioned triplet structure. Vasardani et al. (2013) suggested for a basic PG that ternary relations can be modelled by two edges linking two relatum nodes to the same locatum node. However, this solution is problematic as these edges are not indexed or linked in any way hence can become ambiguous when several of the same non- binary relationship exist for one place. Furthermore, sometimes the sequential order of places can also affect the semantics of non-binary relationships. For non-binary relationships, the task in this chapter is
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 40 placeknowledgemodelling
how to properly model them in order to preserve the original seman- tics and to allow future tracking.
3.2.1.5 Number of Occurrences of Place References and Spatial Relation- ships
The basic PG model does not store the number of times each place reference is used to refer to a place, and thus, the information of which references are more frequently used for a place is lost. Stor- ing the number of occurrences for place references can distinguish between common (popular) names and less-frequently used ones. It also enables analyzing which references are more often used in cer- tain conversational contexts, description themes, or by certain people. The number of occurrences of each relation being used between two places is not recorded in the basic model either, as only one in- stance for each relation can exist between any two nodes. As a result, if two contradicting relations north of and south of between two places have both been stored in a PG, it is impossible to determine which one is more likely to be the true, according to frequencies. By preserving the number of occurrences for each relation, the one that occurs more often can be regarded as a better-agreed upon assertion and, thus, more likely to be true.
3.2.1.6 Conceptualization of Places According to Lynch’s classification of elements of the city (Lynch, 1960), a place from an urban environment can be conceptualized as a node (a strategic spot that is accessible), a path (a channel that affords movement of the observer), a district (an accessible and identifiable area), a landmark (an inaccessible place typically for spatial referenc- ing), or an edge (an inaccessible boundary), as a 0D, 1D, or 2D object. The classification has been adopted in geographic information sci- ence (GIScience), such as for describing the functional spatial structure of urban environments using graphs (Tomko and Winter, 2013). The sense of place emerges as it is functionally different from its surrounding environment and, thus, becomes distinguishable. The functional difference between places is sometimes revealed by place conceptualization in descriptions, and such difference is dependent on the context. The same place can be conceptualized differently in different description contexts or even within the same description (Winter and Freksa, 2012), depending on what information the de- scriptor wishes to convey. For example, a district can be regarded as a 2D container for describing places within it, or being regarded as a 0D landmark for locating other nearby places, either from the same granularity level or not. Capturing the conceptualization of places in descriptions may al- low for better interpretation of the information communicated. For example, the same description ‘the place is to the north of the cam- pus’ can either be interpreted as an external cardinal direction rela- tionship (mapped as north and disjoint) or an internal one (mapped as north and inside), depending on the conceptualization of the relatum,
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 3.2 extending the place graph model 41 i.e., whether the descriptor is regarding the campus as a landmark and describing places nearby, or as a container and describing places within. In the examples, the conceptualization of a place can be re- garded as a variable that affects the mapping of vernacular spatial relationship expressions to formal relations. Without capturing place conceptualizations, the mapping process becomes either risky or un- restrictive.
3.2.1.7 Route and Accessibility Some place descriptions can take the perspective of a route descrip- tion. Route descriptions are often associated with reference directions and accessibility information for navigation purposes. For a triplet in a route description, the accessibility from the relatum to the locatum is usually implied. Accessibility also determines whether the triplet belongs to a part of a route or not. For example, Moncla et al. (2016) use motion expressions to distinguish places that are only seen and places that can actually be reached from hiking descriptions, in order to reconstruct itineraries. In the GeneralizedUpperModel (GUM) on- tology (Bateman et al., 2010), a relationship indicating accessibility is classified as a GeneralizedRoute, and is distinguished from a General- izedLocation which does not belong to a route. Tracking whether places and relationships originated from a route description enables querying of path knowledge for purposes such as navigation support. Moreover, as the number of occurrence of re- lationships is also preserved in an extended PG, it is also possible to identify prominent routes that are described more often by people.
3.2.1.8 Description and Source Contexts Some relations can vary with context. For example, Yao and Thill (2005) identified several contextual variables that determine the choice of qualitative distance relationships, e.g., the current type of activity, and the available mode of transportation. Furthermore, the places referred to in a place description are depending on the purpose of the description. Therefore, information of the theme of a descrip- tion is useful for place analysis. For example, Kim, Vasardani, and Winter (2017a) observed the differences of place occurrences in de- scriptions of four themes: environment, business, travel, and other. In their implementation, places modelled by a basic PG have to be man- ually re-linked to their original descriptions, as the correspondences between places and occurrences in descriptions are not kept when the graphs were constructed. Identifying thematic topics of textual docu- ments can be done using existing techniques such as topic modelling (Adams and Janowicz, 2015; Speriosu et al., 2010). From a database perspective, the metadata of a place description, e.g., its source and time-stamp, should also be preserved. Such knowl- edge can be useful when determining the reliability or time validity of the extracted knowledge. Thus, original speakers and recipients form a facet of context. Individuals may describe the same environment
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 42 placeknowledgemodelling
differently, in terms of the selection of place and place semantics, spa- tial relations, reference frame, and conceptualization. The intention of giving a description to a recipient (or recipients) may also influ- ence how a description will be organized (Wolter and Yousaf, 2018). Human-level factors such as age, gender, ethnicity, and degree of fa- miliarity with the environment have also been identified as being influential on the meaning of the spatial relationships communicated (Yao and Thill, 2005). In addition, the affordance of a place also varies among individuals. For instance, a supermarket may afford for some people shopping, and for others (or at other times) work. Linking descriptions with people allows richer types of queries and analysis to be performed on an extended PG, e.g., what places are more frequently mentioned by whom. As another example, an extended PGD supporting an autonomous vehicle can be used to es- tablish links between passengers’ accounts and the place each one of them calls my home. Such human-related information is often un- available or limited, for example when descriptions are collected from online documents.
3.2.2 The Extended Place Graph Database Model
The extended PGD model is illustrated by the UML diagram shown in Figure 6. The model preserves all the additional information spec- ified so far and is designed to support efficient querying through graph traversal. Each class in the diagram represents a node type in an extended PG, and each relationship represents an edge type. All node types and some edge types are associated with properties. Val- ues of some of the properties will be set to null if the corresponding information is not (or not yet) available. For example, properties such as footprint of a place node cannot be obtained directly from place descriptions, and must be derived using georeferencing techniques. The reason to apply graph database model is that, it opens oppor- tunities for new applications that rely on graph traversal, e.g., QSR regarding a PG as a constrained relational network, landmark identi- fication considering node degree and centrality, measuring place rel- evance by graph distance, or queries by relationship and path. Graph databases also have computational advantages in storing and query- ing (by relationship) richly connected data compared to relational databases, as mentioned in the last chapter already. As an example, if the UML diagram in Figure 6 were modelled by a relational database, queries such as finding the shortest path of spatial relationship con- necting two place nodes would require complex joining operations and may not be done within a reasonable amount of time. An n-plet is an extension of a triplet, and each place reference that occurs in a place description is regarded as being embedded in an n- plet. An n-plet is typically a triplet representing a binary relationship; however, it can also represent a non-binary relationship, e.g., between, around and across, having multiple locata and relata based on the se- quential order of appearance in the description. An n-plet can also
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 3.2 extending the place graph model 43
Figure 6: UML diagram illustrating the extended place graph database model, with seven types of classes (nodes) and nine types of re- lationships (edges). consist of only one locatum without any relatum, as a place reference may not be embedded in any locative expressions in a description, e.g., “Melbourne is a populous city”. Thus, an n-plet must have at least one locatum, and any non-negative number of relata. Examples of modelling binary and non-binary spatial relationships using n-plet are shown in Figure 7.
Figure 7: Examples of modelling a binary spatial relationship west (left) and a non-binary spatial relationship between (right) using n-plet.
Although the proposed database model captures seven types of nodes as conceptual entities, place node is the core of the model, and other nodes are defined in order to store information (includ- ing semantic connections and properties) that is semantically linked to these conceptual entities. In the remaining part of the section, each type of node, edge, and the associated properties are discussed.
3.2.2.1 Place Reference Node A place_reference node represents a reference to a place from an n-plet in a description, either as a locatum or a relatum. Each place reference node must have one and only referred_by incoming edge from a place
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 44 placeknowledgemodelling
node. Between place nodes and place reference nodes are n:1 relation- ships, i.e., a place may be referred to by one or more different place references, while the same place reference may be used to describe different places in different contexts (but modelled by distinct place reference nodes). For example, two references Flinders Street Railway Station (an official place name) and the train station (a non-gazetteered reference) come from conversational contexts where they refer to the same place (Flinders Street Railway Station). In a different context, the reference the train station may refer to another train station. A place reference node, when created, is by default linked to a new place node instance through a referred_by edge. A merging algorithm (Kim, Vasardani, and Winter, 2017b) can then modify the correspon- dence by removing the newly created edge and place node instance, and establishing another referred_by edge between the place reference node and a pre-existing place node, if it is determined so. Since place references are embedded in n-plets extracted from place descriptions, each place reference node has one and only outgo- ing edge in to an n_plet node. An in edge has two mandatory prop- erties: pos and as. The value of as can either be locatum or relatum, representing whether the place reference is corresponding to the lo- catum or the relatum of the n-plet. The property pos is a positive integer denoting the index of the occurrence of the place reference, as it is possible that an n-plet has multiple locata or relata. For a triplet, the value of pos is 1 for either of the two place reference nodes it links to. Thus, a place reference node is defined by Axiom 1 below:
Place_reference v ∃referred_by−.Place u ∃in.N_plet (1)
A place reference node has six properties: place_reference, concep- tualization, place_type, equipment, characteristic, and affordance. Table 1 provides details for each property. Among the properties only the first one place_reference is mandatory. The value of conceptualization is one of the categories based on Lynch’s classification: node, path, dis- trict, landmark, or edge. Values for the remaining four properties are unrestricted, and some examples have been given in Section 3.2.1.1. The data type of these four properties is string list, as multiple values of each of these properties can be described. These properties are not stored under place nodes in order to preserve the context of where and by whom these values are given.
3.2.2.2 N-Plet Node An n-plet node is defined by Axiom 2 below. Other than in as already explained, each n-plet node has an outgoing edge from, denoting the description from which this n-plet is extracted. The edge from has the same property pos as in, showing the sequential order of appearance of an n-plet in the description. An n-plet node can have one or more
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 3.2 extending the place graph model 45
Table 1: Properties of a place reference node.
Property Data type Details
place_reference string a reference referring to a place, e.g., a place name Paris conceptualization string one from {node, path, district, landmark, edge} place_type string list a list of place types, e.g., park equipment string list a list of nouns, e.g., food characteristic string list a list of adjective or descriptive phrases, e.g., old affordance string list a list of activities either as verbs or verb phrases, e.g., eating in and map edges, depending on the number of locata and relata, and the number of mapped formal spatial relations mapped respectively.
N_plet v ∃in−.Place_reference u ∃from.Description (2) u∃map.Spatial_relation
An n-plet node has two properties: spatial_relation_expression and reference_frame. Table 2 provides details for each property. The first one stores the original spatial relationship expression used for the n- plet in the description. In a basic PG, such expressions are formalized by a controlled vocabulary before graph construction (Kim, Vasar- dani, and Winter, 2016), yet it is quite often that the same spatial relationship expression can be mapped to different formal relations depending on the context. Therefore, in an extended graph, the orig- inal spatial relationship expressions are kept, and the mapped rela- tionships will be stored separately as spatial_relation nodes linked by outgoing edges map from n-plet nodes.
Table 2: Properties of an nplet node.
Property Data type Details
spatial_relation_expression string the original expression of a spatial re- lationship, e.g., left-hand side reference_frame string one from {intrinsic, relative, unknown}
If the spatial relation expression of an n-plet is mapped to a rela- tive direction relation, the value for the property reference_frame can either be intrinsic, relative, or unknown (undetermined). The intrinsic value means the relative direction is based on the intrinsic direction of the relatum (e.g., ‘in front of the building’), while the relative value means a non-intrinsic reference direction is adopted. If the value rel- ative is used, the n-plet node will have an additional outgoing edge has_reference_direction to a place reference node referred in the dis- course, anchoring the reference direction used for the n-plet. An edge of has_reference_direction has a property as, and the value is one from {front, back, left, right}. An example is given in Figure 8, modelled from the following description:
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 46 placeknowledgemodelling
“... coming from the Main South Entry, the Baillieu Li- brary will be on the left hand side of the South Lawn ...”
Figure 8: An example of modelling a relative direction relationship using the extended place graph model.
In the example, the interpretation of the relative direction re- lation
3.2.2.3 Place Node
Each place node represents a place. In an extended PG, a place is iden- tified from one or more place descriptions by place references em- bedded in n-plets. A place node does not have any place references stored; however, all the references used for referring to it (as well as the number of occurrences for each reference) can be obtained easily from all the place reference nodes it is connected to through outgoing referred_by edges. A place node is defined by Axiom 3:
Place v ∃referred_by.Place_reference (3)
A place node has three derived properties, as shown in Table 3. The value of footprint represents the location of the place, and can be either a point, a polyline, a polygon, or an approximate location region (ALR) that will be introduced in Chapter 5. An ALR is a prob- abilistic field derived using spatial relation search space models for georeferencing places using their described spatial relations to other places. The value of property type denotes the data type of the foot- print, e.g., polygon. The property spatial_granularity is a classification of the spatial granularity of the place based on the categories pro- posed by Richter et al. (2012): {furniture, room, building, street, district, city, country}. The footprint value type is set to GeoJSON as the de- signed database model will be implemented by a graph database, which typically does not allow storing probabilistic surfaces. For an
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 3.2 extending the place graph model 47
ALR, the corresponding footprint value is a polygon of its crisp bound- ary given a threshold.
Table 3: Properties of a place node.
Property Data type Details
footprint GeoJSON as a point, polyline, a polygon, or an ALR footprint_type string one from {point, polyline, polygon, ALR} spatial_granularity string one from {furniture, room, building, street, dis- trict, city, country}
3.2.2.4 Route Node Places referred to as part of a route are grouped by linking their cor- responding place nodes to a route node through part_of edges. The property pos of a part_of edge records the position of the place refer- ence in the route by sequential order of appearance, and the value is a positive integer.
3.2.2.5 Spatial Relation Node Each spatial_relation node represents a formal spatial relation. Unlike a value of the property spatial_relation_expression stored in an n-plet node, which can be expressed in flexible ways, formal relations are from a controlled vocabulary. Binary formal relations from four fami- lies are considered, as listed in Table 4. The vocabulary of non-binary relations in not restricted, since currently non-binary relations are less studied in the literature. Table 5 provides details for each property. The value of property family is from one of the five spatial relation families: {cardinal_direction, qualitative_distance, relative_direction, topo- logical, non-binary}. The property relation stores the name of a relation from one of the five families.
Table 4: Formal binary spatial relations from different families.
Spatial relation family Spatial relation type
Cardinal direction north, south, east, west, northeast, southeast, northwest, southwest Qualitative distance near, far Relative direction front, back, left, right, left front, right front, left back, right back Topology inside, covered by, overlap, meet, disjoint, cover, contain, equal
Mapping between spatial relation expressions and formal rela- tions is a m:n relationship. A spatial_relation_expression value can be mapped to one or more formal relations from single or multiple fam- ilies, and different spatial relation expressions could be mapped to the same formal relation. The mapping process is context-dependent. For example, a spatial relation expression ‘north’ can be mapped to either north, disjoint (external north) or north, inside (internal north),
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 48 placeknowledgemodelling
Table 5: Properties of a spatial relation node.
Property Data type Details
relation string a spatial relation, e.g., near family string one from {cardinal_direction, qualitative_distance, rela- tive_direction, topological, non-binary}
depending on the original expression and place conceptualization. Compared to a basic PG, the extended model supports more flexible and context-aware reasoning of spatial relations.
3.2.2.6 Description Node A description node represents a description document as a single dis- course. It is used to store the global-level context variables of the descriptions from which n-plets were extracted. The four properties of a description node, i.e., theme, transportation_mode, source, and times- tamp, have already been explained in Section 3.2.1.8. A description node has at least one instance of an incoming edge from from an n-plet node. The property pos of from is the position of the linked n-plet in the description by appearance, and the value is a positive integer. The property spatial_context is a derived one, rep- resenting the geographic extent a description is embedded in, using the approach developed in Chapter 5. For example, if the extracted places are landmarks in the Melbourne CBD as a suburb, the con- text of the original description is likely to be about Melbourne CBD. Finally, a description node can also have outgoing edges created_by and given_to to a user node, if such information is available. Table 6 provides details for each property.
Table 6: Properties of a description node.
Property Data type Details
theme string a keyword showing the topic of the descrip- tion, e.g., tourism transportation_mode string a keyword showing the transport method, e.g., walking source string a keyword showing the source of the de- scription, e.g., Wikipedia timestamp string a timestamp e.g., 2018-5-28
3.2.2.7 User Node A user node either represents a descriptor (connected by a created_by edge from a description node) or a recipient (connected by a given_to edge). The same user node can be connected to multiple description edges by either roles. The only property info of a person node is not restricted in this chapter, as what information is useful for the appli- cation of an extended PG is domain- and task-dependent. Examples
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 3.3 implementation 49
have been given. The property value is defined in the format of JSON as key-value pairs.
3.3 implementation
This section describes the place description datasets used in this the- sis, the implementation of the extended PG model, as well as a com- prehensive example demonstrating how a description can be mod- elled using the proposed model. The constructed PGDs in this section will be used in the later three chapters for reasoning, georeferencing, and querying tasks.
3.3.1 Data Overview and Graph Construction
The extended PG model has been implemented using Neo4j graph database1 and Python with a database management system interface. The system also integrates functions that will be explained in the later chapters (i.e., reasoning, georeferencing, and querying) to provide a complete processing chain. The system is additionally able to per- form automatic graph creation from JSON input, graph visualization, and place mapping. A complete demonstration of the system will be provided in Chapter 7. Two sets of place descriptions are used in this thesis. The first one contains 42 descriptions submitted by graduate students about the University of Melbourne campus. It has richer spatial relationship connections among places and more focused spatial coverage com- pared to the second dataset. The second dataset was harvested from web texts for places around and inside the area of the Greater Mel- bourne, Australia (Kim, Vasardani, and Winter, 2015). The sources for the Melbourne dataset include: WikiMapia as a collaborative map- ping platform with user generated place descriptions; Wikipedia ar- ticles with descriptions of places; business sites or official sites with descriptions related to locations such as of companies, shops, and restaurants; and blogs with descriptions focused on individual inter- ests such as tourism. The types of geographic features in the datasets vary from fine-grained local points-of-interests to large geographic features such as nature reserves. Descriptions from certain sources are more likely to include certain types of places. For instance, busi- ness sites typically focus on urban contexts, while tourism articles may be from either urban contexts or natural environment contexts. It is also observed that places from urban contexts are noticeably finer in granularity and more frequent than places as natural geographic features from the datasets. The two datasets cover more than 3000 distinct places. Note that a place can be referred to by multiple place references in different descriptions, or even within the same descrip- tion. In summary, the datasets cover descriptions with different sizes, granularity, and place density. An example description is shown be- low:
1 https://neo4j.com/
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 50 placeknowledgemodelling
“... If you go into the Old Quad, you will reach a square courtyard and at the back of the courtyard. You can either turn left to go to the Arts Faculty Building, or turn right into the John Medley Building and Wilson Hall. Raymond Priestly Building is the open aired ground area which is in front of Wilson Hall that is adjacent to it. Towards North, which is when you turn left when exiting the Old Quad, you will see Union House where there are shops selling foods. If you continue walk along the road on the right side where you’re facing Union House, you can see the Beaurepaire and Swimming Pool. There will also be a sport tracks and university oval behind it ...”
Place name recognition is outside the scope of this thesis, and there- fore triplets extracted by a previously-developed parser from each of the description dataset are used (Liu, Vasardani, and Baldwin, 2014). Non-binary n-plets are not considered in the implementation, due to the limitation of the used parser. The parser applies sentence and word tokenisation, part-of-speech (POS) tagging, as well as full-text chunk parsing. It trains a conditional random fields (CRF) model on a manually annotated corpus. To enhance its performance, it also ex- ploits external resources including gazetteers. Since place name ex- traction precision and recall are irrelevant in this thesis, all place names had been cleaned manually beforehand, and incorrectly ex- tracted noises such as person names are stripped-off. Place references referring to the same places have been identified using the technique introduced before (Kim, Vasardani, and Winter, 2017b). For the experiments in the later chapters, the reference frame and direction information of triplets with relative direction relations is useful as well. However, such information cannot be obtained using the parser. Therefore, a manual annotation process by two graduate students is performed, and only the annotations that both students agree with are accepted. To minimize the influence of pre-existing lo- cal knowledge on the annotation process, all place references were re- placed by five place types: building, spot (any place finer than a build- ing), area, alley, and street. For example, the first building name that occurs in a description is anonymized as Building_1. The task of the student was to assign each relative direction relationship with three property values: reference frame (intrinsic or relative), the anonymized reference of the place indicating the reference direction (e.g., ‘Main South Entry’ in Figure 8), and the reference direction (value of prop- erty as for edge has_reference_direction).
3.3.2 Demonstrative example
An example description from the experimental dataset is shown be- low. In the description, references to places and spatial relationship expressions are highlighted using bold and italic font receptively. Other non-spatial information, such as place equipment and place characteristics, is highlighted by underlines. The example illustrates
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 3.3 implementation 51
all types of information identified in Section 3.2.1. Note that due to the lack of available parsers and description metadata, currently some information, such as user information, place characteristics, and trans- portation modes, cannot be obtained automatically, and the parsed file below contains such information that is manually filled-in. The example is for illustration purpose only, and such unavailable infor- mation will not be used in the experiments in later chapters.
“South Lawn is the major reference point which is situ- ated in about the middle of the campus. Coming from the Main South Entry, the Baillieu Library will be on the left hand side of the South Lawn. To the north of this you have the Old Quad (really old English style building). If you want food and are currently on South Lawn go through the Old Quad to the north and keeping heading north until you get to a Union House.”
The description is first parsed into a JSON file with all the extracted information, as shown below:
1 {"descriptions":[
{"did": 1, "theme":"University of Melbourne campus description", "transportation mode":"walking", 6 "descriptor": {"uid": 1, "identity":"University of Melbourne student"}, "recipient": {"uid": 2, 11 "identity":"University of Melbourne student"},
"n plets": [{"nid": 1, "locatum reference":"South Lawn", 16 "locatum characteristic":"the major reference point", "locatum conceptualization":"node", "spatial relation expression":"about the middle of", "relatum reference":"campus", "relatum conceptualization":"district", 21 "relation map":"inside"},
{"nid": 2, "locatum reference":"Baillieu Library", "locatum conceptualization":"node", 26 "spatial relation expression":"on the left hand side", "relatum reference":"South Lawn", "relatum conceptualization":"node", "reference frame":"relative", "reference direction": ["Main South Entry","back"], 31 "relation map":"left"},
{"nid": 3, "locatum reference":"Old Quad", "locatum type":"building", 36 "locatum characteristic": ["old","English style"], "locatum conceptualization":"node", "spatial relation expression":"to the north", "relatum reference":"South Lawn", "relatum conceptualization":"node", 41 "relation map":"north"},
{"nid": 4,
[ April 29, 2019 at 20:02 - classicthesis version 5 ] 52 placeknowledgemodelling
"route_id": 1, "locatum reference":"Union House", 46 "locatum equipment":"food", "locatum conceptualization":"node", "spatial relation expression":"go through", "relatum reference": ["Old Quad","South Lawn"], "relatum conceptualization": ["node","node"], 51 "relation map":"through"},
{"nid": 5, "route_id": 1, "locatum reference":"Union House", 56 "locatum conceptualization":"node", "spatial relation expression":"to the north and keeping heading north", "relatum reference":"Old Quad", "relatum conceptualization":"node", "relation map":"north"}] 61 } ]} As there is only one description, a description node with two user nodes will be created, as shown in Figure 9. Then, five n-plet nodes will be created and linked by from edges to the description node, as shown in Figure 10. Next, place reference nodes, mapped spatial re- lation nodes, and a route node (if the n-plet is part of a route) will be created and linked to each n-plet node, as shown in Figure 11. Last, a place node will be created for each place reference node. The node merging approach mentioned previously is then used to link places that are referred by different place references from different n-plet, resulting in multiple place reference nodes linked to the same place node, as shown in Figure 12. Figure 13 shows the resulting extended PG of the whole place description generated using Neo4j.
Figure 9: Creation of description and user nodes.
3.4 chapter discussion and summary
This chapter proposes an extended PG model for information ex- tracted from place descriptions, including references to places, spa- tial relationships, place properties and various contextual knowledge. The model is a revision and an extension of a previously proposed basic PG model in order to overcome its limitations, in terms of the types of information that can be modelled, as well as they way they are modelled. Eight types of information that are embedded in place descriptions and not captured in the basic PG model are
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 3.4 chapter discussion and summary 53
Figure 10: Creation of n-plet nodes.
Figure 11: Creation of place reference, spatial relation, and route nodes.
Figure 12: Linking different place reference nodes to the same place node through node merging. identified: place semantics and characteristics, places and spatial rela- tions from the same discourse, as well as their sequential order of ap- pearance, reference frame, non-binary relationships, co-occurrence of place references and spatial relations, place conceptualization, route and accessibility, and description context and source context. A graph
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 54 placeknowledgemodelling
Figure 13: The resulting extended place graph database of the place descrip- tion example. Gray: description; Pink: user; Green: n-plet; Blue: (mapped) spatial relation; Red: route; Yellow: place reference; Pur- ple: place.
database model is then designed for modelling these types of in- formation, which includes seven types of nodes as conceptual enti- ties: place reference, n-plet, place, route, spatial relation, description, and user, as well as nine types of edges representing the semantic connec- tions among these entities. Both nodes and edges are associated with properties. The model allows convenient and efficient query through graph traversal, which is ideal for tasks in the later chapters of this thesis. The model is implemented using a commercial graph database platform with two place description datasets for later experiments, and a database management system interface capable of various PG operations has been developed as well. The extended PG model captures the five core concepts of spatial information proposed by Kuhn (Kuhn, 2012): location, field, object, net- work, and event. The graph stores the location of a place, possibly as a field (approximate location region). Different places are modelled as node objects in a PG, characterized by information including place references and semantics. An extended PG also forms a network by rep- resenting the links not only between places, but also between places and descriptions, as well as between places and people. Such links can be strengthened given additional descriptions, as their times of co- occurrence are captured as well. Finally, an event involves aspects of people, time, location, and activity, and these aspects are covered by node properties info (of user nodes), timestamp (of description nodes), footprint (of place nodes), and affordance (of place reference nodes) re- spectively.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 3.4 chapter discussion and summary 55
The model is not limited by language, although the implementa- tion of the model in this thesis uses datasets in English. Spatial re- lationships are typically expressed in NL as prepositions, as phrases, or implicitly. Once spatial relationship expressions are extracted by a parser and mapped to formal spatial relations, they can be modelled by the proposed place graph database model in their own language. The uses of spatial expressions are language-dependent, and there- fore different parsers must be used for place descriptions in various languages. Other non-spatial information such as place affordance and characteristics can be modelled as well and is not limited by lan- guage. The extensions by the model were identified regarding the needs of this thesis for richer knowledge. It is anticipated that additional fu- ture extensions are possible considering application scenarios, since nodes, edges, as well as their properties in the proposed model are extensible. Some of the identified information, however, requires new techniques or parsers to obtain this automatically, which are not readily available. Although natural language processing (NLP) is be- yond the scope of this thesis, it is necessary for the automation of extracting several types of information identified. Spatial language, although comprehensively studied in linguistic and spatial cogni- tion, is relatively ignored in the community of NLP. Machine learn- ing seems promising for spatial language processing and has been used previously for extracting place references as well as spatial re- lationships. For example, it can be used for extracting motion verbs indicating routes, phrases indicating human activities at places, or ex- pressions of non-binary relationships. Techniques being able to per- form information extraction (IE) and POS tagging are preferable for these types of problems, such as hidden Markov model (HMM) or CRF. It is also promising to apply deep learning, which has pushed the state-of-the-art of many machine learning problems, with neu- ral network models that were used for IE and POS tagging problems before, such as recurrent neural networks (RNN) or long short-term memory (LSTM).
[ April 29, 2019 at 20:02 – classicthesis version 5 ] [ April 29, 2019 at 20:02 – classicthesis version 5 ] QUALITATIVESPATIALREASONING 4
A place graph database (PGD) captures the qualitative spatial rela- tionship knowledge between places as expressed in verbal or written place descriptions. Such knowledge could be used for qualitative spa- tial reasoning (QSR) tasks, such as inferring new spatial relationships based on existing ones, or identifying logical contradictions among the stored spatial relationships. QSR is well-studied in the field of ar- tificial intelligence with formally defined spatial relation reasoning models typically for single family, and these models have not been used for reasoning with collective human spatial knowledge before. A PGD captures spatial relationships from different families, and the variety and flexible use of natural language (NL) may not apply these formal models. This chapter focuses on the task of relational consistency mainte- nance in a PGD from a pragmatic rather than logical perspective. A flexible reasoning framework is developed based on relational com- position rules as well as a path-based consistency checking algorithm. Building on the established field of QSR, four types of spatial rela- tionship families are considered: cardinal direction, relative direction, qualitative distance, and topology. The experiment results show that the approach is reliable and robust in maintaining relational consis- tency during database transactions within a limited path length. This chapter is based on content from previously published papers, including a reasoning framework (Chen, Vasardani, and Winter, 2015) and an extension of reasoning with relative direction relations us- ing reference frame knowledge captured in a PGD (Chen et al., 2018). The contributions of the corresponding (first) author include prob- lem conceptualization, literature review, methodology, implementa- tion, experiments, and paper writing.
4.1 introduction
People use various qualitative spatial relations to describe the rela- tive locations of places within some environments, such as direction (e.g., left, north), distance (e.g, near), and topology (e.g., inside). The PGD model presented in Chapter 3 allows storing of such relational knowledge extracted from place descriptions. The goal of QSR is to allow a machine to represent and reason with spatial entities with- out using traditional quantitative techniques (Cohn and Renz, 2008). Therefore, as a knowledge base of qualitative spatial relationships, a PGD is expected to be capable of QSR tasks. Specifically, this chapter focuses on the task of maintaining relational consistency in a PGD.A human is able to detect logical contradictions among spatial relation- ships, at least for simple scenarios. In order to maintain and query a large PGD, a mechanism to maintain relational consistency of the
57
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 58 qualitativespatialreasoning
knowledge stored during database transactions (e.g., creation or up- date) is necessary. The consistency of the knowledge is also crucial for the later georeferencing and querying tasks. Qualitative spatial relationship models have been developed in the field of QSR for applications such as robot navigation, which seems applicable for the the consistency reasoning task in this chapter. How- ever, NL spatial relationship expressions are not necessarily commit- ting to any of these logical models. For example, cardinal directions have been formalized by different models such as linear, conic or half- plane ones, while the variety and flexible use of NL prepositions does neither convey which one to apply nor does it guarantee strict co- herence (Klippel and Montello, 2007; Klippel et al., 2004). Therefore, models from the QSR community may not be flexible enough for ac- commodating NL spatial relationship expressions and thus, have to be modified for the task of this chapter. Moreover, a PGD contains com- binations of qualitative spatial relations of multiple families, while frameworks developed in QSR are typically realized for single family, such as topological, directional, or distance relations. Furthermore, a mechanism to apply QSR models for consistency maintenance consid- ering the structure of a PGD and computational feasibility has to be developed. This chapter contributes a flexible framework for relational consis- tency maintenance in a PGD. For this purpose this chapter studies the relationships and calculi in question, suggests novel composition rules as reasoning constraints and a path-based algorithm for con- sistency checking, and tests them. The framework ensures local con- sistency (i.e., consistency maintained considering only limited path length) instead of global. However, experiment results demonstrate that the framework is reliable in identifying and flagging inconsis- tent relationships (but not deciding which is true) even with prag- matically limited path length. For implementation binary spatial re- lations from four families are considered: cardinal direction, relative direction, qualitative distance, and topology, since these families have been well-studied in the field of QSR. The framework first deals with spatial relations from each family, and it is then extended to support cross family reasoning. The chapter provides a first step into reason- ing with collective human qualitative spatial relationship knowledge using a graph-based model. The remainder of this chapter is structured as follows. The reason- ing framework, including an illustration of the problem, relational composition tables, as well as a consistency checking algorithm, are explained in Section 4.2. Experiments for testing the framework and the corresponding results in both single family and cross family sce- narios, are provided in Section 4.3. Reasoning considering paths of different lengths is also tested and compared. Section 4.4 presents a discussion for both the developed framework as well as experiment results and concludes the chapter.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 4.2 consistency reasoning in a place graph database 59
4.2 consistency reasoning in a place graph database
A PGD can be set up with multiple internal consistency constraints. Structural ones such as that no node should be disconnected, and no edge should be open-ended, can easily be enforced as database trans- action rules. This chapter instead focuses on semantic and logical consistency rules, ensuring that reasoning along spatial relationships should not lead to contradictions. This section starts with an illustra- tion of the problem. Then, a framework for the problem is presented.
4.2.1 Problem Elucidation
The task of this chapter is divided into two sub-problems: relation in- ference and consistency maintenance. An example illustrating the two problems is shown in Figure 14. The left figure shows the spatial configuration of three spatial objects, with two spatial relationships describing the configuration given in the right figure, that
Figure 14: An illustrative example of relational inference and consistency reasoning problems. The spatial configuration of three spatial ob- jects (left), and two spatial relationships describing the configura- tion (right).
Relational inference can be regarded as a sub-problem of consis- tency maintenance. As illustrated by Figure 14, in order to determine whether a new spatial relationship between A and C occurred in a transaction is consistent with existing knowledge, one solution is to first determine what are the possible spatial relationships between A and C through inferring based on existing relational knowledge, and then check if the new spatial relationship is a subset of the inferred possible ones. Therefore, the task of this chapter can be refined as: maintaining relational consistency in a PGD considering inferred rela- tional knowledge.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 60 qualitativespatialreasoning
More specifically, with each creation or update operation of a PGD, consistency with the already stored knowledge can be checked. Dele- tions of entries leave a relationally consistent database always in a relationally consistent state. The simplest form of place description, a single spatial relationship, is always consistent as a singly connected path. If one of, or both the locatum and the relatum do not exist in the database, then adding a new spatial relationship will not cause re- lational inconsistency. Therefore, only additions forming cycles have to be watched for inconsistency, and with addition of knowledge the following cases can be distinguished:
• Between two nodes a relationship is added in the same direction as an existing one (e.g., a new relationship
• Between two nodes a relationship is added in the opposite di- rection of an existing one (e.g., a new relationship
• Between two not yet connected nodes a relationship is added. This is again a case of relation composition, over paths in the database of a length of at least two edges and forms a cycle of length 3 or more, e.g., a new relationship
If the new knowledge is incompatible, the database should flag an inconsistency and ask for human intervention, accepting that prag- matics may overrule. Two definitions are given:
Definition 1 (Consistent spatial relationship) A spatial relationship is re- lationally consistent if no contradicting spatial relationships can be inferred from the place graph database where it is stored.
Definition 2 (Consistent place graph database) A place graph database is relationally consistent if, at its current state, it is free from inconsistent spatial relationships.
Therefore, the purpose of this chapter is to develop a mechanism to ensure a PGD is consistent at any state after transactions. Next, a rea- soning mechanism for maintaining relational consistency that works for all of the three cases is introduced. Short names of relations are used in the following sections. Topological relations (disjoint, meet, equal, covered by, contain, overlap) are denoted by {d, m, e, i, cB, ct, cv, o}, qualitative distances (near, middle, far) by {n, m, f}, cardinal direction relations (north, south, east, west, northeast, southeast, northwest, south- west) by {N, S, E, W, NE, SE, NW, SW}, and relative direction relations (front, right, back, left, left front, right front, left back, right back) by {F, R, B, L, LF, RF, LB, RB}.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 4.2 consistency reasoning in a place graph database 61
4.2.2 Single-Family Relational Inference
A compositional inference is a deduction based on two relationships rA,B and rB,C, which results in rA,C. Sometimes the inference result i is a set of possible relations SA,C = {rA,C}. The symbol ◦ is used to de- note a composition operation between two relationships. A table can be constructed to represent the composition results SA,C for spatial relations in a family. For the composition of n possible relationships in a family, an n ∗ n composition table can be defined. When compos- ing more than two relations, for instance rA,B ◦ rB,C ◦ rC,D, the result will be all possible relationships between A and D, or SA,D. Such com- position over long sequences of places and spatial relationships can be reduced to binary cases, i.e., compositing two spatial relations at a single time. There are three general laws for spatial relation composition as shown below. However, the associative or commutative laws do not necessarily apply, i.e., the result between r1 ◦ r2 and r2 ◦ r1, or those between (r1 ◦ r2) ◦ r3 and r1 ◦ (r2 ◦ r3) could be different for some spa- tial relationship families. Specifications of applying conditions will be provided in the following subsections.
• Commutative law: r1 ◦ r2 = r2 ◦ r1
• Associativity law: (r1 ◦ r2) ◦ r3 = r1 ◦ (r2 ◦ r3)
• Distributive law: (r1 ∨ r2) ◦ r3 = (r1 ◦ r3) ∨ (r2 ◦ r3)
Composition tables have been defined in the community of QSR for different models. In the remaining part of this section, composition tables that are applied for each of the four family considered are in- troduced. cardinal direction relation Frank (1992) suggested reason- ing models with cardinal directions relationships, as shown in Fig- ure 15, including a cone model (a), a half-plane model (b), and a neutral zone model (c). The neutral zone model with its composition table is applied in this chapter. The size of the neutral zone is defined by the relatum, while the locatum belongs to one of the eight zones (NW, N, NE, E, SE, S, SW and W).
Figure 15: Qualitative spatial reasoning models of cardinal direction rela- tions proposed by Frank (1992), including a cone model (a), a half-plane model (b), and a neutral zone model (c).
Table 7 shows the corresponding composition table of the neutral zone model (Frank, 1992). any in the table means that any of the eight
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 62 qualitativespatialreasoning
relationships is possible. However, the semantics of cardinal direc- tions in NL are more flexible than this logical model suggests. Fig- ure 16 shows a failure case of the composition table of the neutral zone model when applying to the knowledge stored in a PGD. The left figure shows two spatial relationships among three places A, B, and C. According to Table 7, the inferred cardinal direction relation- ship between C and A will be N ◦ NE = NE. However, as shown in the right figure, a spatial configuration exists that allows C to be located to the northwest of A while not violating the two spatial relationships shown in the left figure. This is because according to the neutral zone model, C should be described as having a NW relation to B and B should be described as NE to A, while in NL people may use spatial relations more flexibly, such as the ones shown in the left figure.
Table 7: Frank’s composition table for the neutral zone model for cardinal direction relations Frank (1992). N NE E SE S SW W NW N N NE NE E any W NW N NE NE NE NE E E any N N E NE NE E SE SE S any N SE E E SE SE SE S S any S any E SE SE S SW SW W SW W any S S SW SW SW W W NW N any S SW SW W NW NW NW N N any W W NW NW
Figure 16: A failure case of the proposed composition table for the neutral zone model when applied to natural language spatial relationship expressions.
The composition table has therefore been modified in this thesis, as shown in Table 8. An extended cardinal direction relation set (ECD) is additionally defined as a set of cardinal direction relations that are not contradicting with a given cardinal direction relation in NL, i.e., for an element in an ECD for a spatial relation, there exists a spatial configuration that both the element and the spatial relation can be used to describe the configuration. The ECD for each cardinal direction relation is shown in Table 9. ECD is used in this way: the possible spatial relationships between two places is composited over a sequence of places and spatial relationships using Table 8, and any new spatial relationships between these two places that are elements
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 4.2 consistency reasoning in a place graph database 63 of the ECDs of the composited possible spatial relationships will be regarded as consistent. Therefore, for the example shown in Figure 16, the inferred spatial relationship between C and A will be N ◦ NE = N, and any new spatial relationships between C and A that belongs to the ECD of N, i.e., {W, NW, N, NE, E}, will be regarded as consistent.
Table 8: A modified composition table for cardinal direction relationships. N NE E SE S SW W NW N N N any any any any any N NE N NE E E any any any N E any E E E any any any any SE any E E SE S S any any S any any any S S S any any SW any any any S S SW W W W any any any any any W W W NW N N any any any W W NW
Table 9: Extended cardinal direction relation sets of cardinal direction rela- tions.
Original relation ECD
N W, NW, N, NE, E NE NW, N, NE, E, SE E N, NE, E, SE, S SE NE, E, SE, S, SW S E, SE, S, SW, W SW SE, S, SW, W, NW W S, SW, W, NW, N NW SW, W, NW, N, NE
ECD provides a tolerant mechanism that first allows all possible cardinal direction relations to be stored that are considered to be consistent. Then, when additional knowledge is stored as the PGD get populated or updated, the constraint may be refined, and a new input spatial relationship will need to be consistent with all exist- ing knowledge in order to be regarded as consistent. An example is shown in Figure 17. Based on the knowledge shown in the left fig- ure, the possible cardinal direction relationships between C and A are {N, NE, E, SE, S} through composition, while based on the knowledge in the right figure the result is {W, NW, N, NE, E}. Assuming knowl- edge from both figures is stored in the same PGD, the possible cardinal direction relationships between C and A are refined to {N, NE, E} as the intersection of the two sets. Details of the formal algorithm for consistency maintenance will be provided in Section 4.2.3. qualitative distance relation For qualitative distance rela- tionships Frank (1992) proposed two-step (near, far), three-step (near, middle, far) and multi-step systems for reasoning. These models con- sider qualitative distance as mappings into intervals that form a parti-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 64 qualitativespatialreasoning
Figure 17: Refinement through ECD intersection with additional spatial re- lation knowledge.
tion of the positive real numbers, such as (for a three-step system) near → [0, 1); middle → [1, 3); far → [3, ), in order to allow interpretation. An illustration is shown in Figure 18 (left). In a linear space such re- lationships can be added. The resulting∞ interval will then be mapped back to the corresponding symbols (near, middle, far). However, quali- tative distance relationships in a PGD are from two-dimensional space, where the triangle inequality holds. Figure 18 (right) shows for exam- ple a situation where C is far from B and B is far from A, while A and C can be in any relationship, including near. The composition table provided by Frank (1992), in comparison, defines f ◦ f to be f.
Figure 18: A qualitative distance relationship model for reasoning (left), and the limitation of the model when considering triangle inequality (right).
Therefore, the three-step composition table is modified and applied in this chapter. Table 10 shows the corresponding composition table, using the distance computations of Algorithm 1. In the algorithm, ra and rb are from the set of {n, m, f}. Instead of directly adding the lower and upper bound of the intervals of the relations to be compos- ited (e.g., compositing f → [3, ) ◦ f → [3, ) results in the interval of [0, ) and therefore is mapped to any). The algorithm considers triangle inequality in non-linear∞ situations. ∞ ∞ Table 10: Modified composition table of qualitative distance relations based on Algorithm 1.
n → [0, 1) m → [1, 3) f → [3, )
n → [0, 1) n, m → [0, 2) any → [0, 4) m, f → [2, ) ∞ m → [1, 3) any → [0, 4) any → [0, 6) any → [0, ) f → [3, ) m, f → [2, ) any → [0, ) any → [0, ∞) ∞ The intervals∞ in Table 10 are∞ directly borrowed∞ from∞ the composi- tion table proposed by Frank (1992). Note that qualitative distance
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 4.2 consistency reasoning in a place graph database 65
Algorithm 1 Distances for qualitative distance interval composition.
1: ra = [xa, ya), rb = [xb, yb) 2: yc = ya + yb 3: if xb > xa then 4: xc = |xb − ya| 5: else 6: xc = |xa − yb| 7: end if 8: return [xc, yc) relations are flexibly used by people in NL, and such relations are often highly vague and context-dependent by various factors. For ex- ample, near in a coarse-granularity spatial context may correspond to a geometrical distance that is regarded as far in a finer-granularity context. Therefore, the proposed composition table can only assumes a limited context. topological relation For the topological relation family, the composition table for the 4-intersection model proposed by Egenhofer (1991) is applied, shown in Table 11. Again, any means the universal relationship: any topological relationship is possible. The commuta- tive law does not apply for the composition operation between topo- logical relationships. For instance, d ◦ i = {d, m, i, cB, o} while i ◦ d = d.
Table 11: Composition table of topological relations. d m e i cB ct cv o d any d,m,i, d d,m,i, d,m,i, d d d,m,i, cB,o cB,o cB,o cB,o m d,m,ct, d,m,e, m i,cB,o m,i, d d,m d,m,i, cv,o cB,cv,o cB,o cB,o e d m e i cB ct cv o i d d i i i any d,m,i, d,m,i, cB,o cB,o cB d d,m cB i i,cB d,m,ct, d,m,e, d,m,i, cv,o cB,cv,o cB,o ct d,m,ct, ct,cv,o ct e,i,cB, ct,cv,o ct ct ct,cv,o cv,o ct,cv,o cv d,m,ct, m,ct, cv i,cB,o e,cB, ct ct,cv ct,cv,o cv,o cv,o cv,o o d,m,ct, d,m,ct, o i,cB,o i,cB,o d,m,ct, d,m,ct, any cv,o cv,o cv,o cv,o
relative direction relation Relative direction relationships require a reference direction in order to be anchored, while all car- dinal direction relationships are under the same, absolute reference frame. Thus, two spatial relationships
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 66 qualitativespatialreasoning
reference direction. In a PGD, reference directions of relative direction relationships are captured. While the formal reasoning frameworks for relative direction rela- tions in QSR typically assume the reference direction is known, the proposed place graph (PG) model allows storing the reference direc- tion of a relative direction relationship either by pointing to another place from the discourse, or as an intrinsic one. Reference direction transformation rules are applied for relations from the first case, since the intrinsic directions of places in the second case cannot be an- chored without further knowledge. An illustration of the reasoning process is provided in Figure 19. Given the knowledge in (a), a consis- tent relationship can be inferred as shown in (b), through transforma- tion of the reference direction. Given another relationship
Figure 19: Consistency reasoning through reference direction transforma- tion. An existing relationship (a); another new relationship that is consistent with the existing one, determined by reference direc- tion transformation (b); another relationship that is inconsistent with the previous two (c).
The reference direction transformation process is based on a com- prehensive list of rules, and each relative direction relationship with a locatum, a relatum, and a reference place, e.g., A, B, and C in Fig- ure 19 (a), can be used to infer five other relative direction spatial relationships, including the one shown in Figure 19 (b). This is be- cause any one of A, B, and C can become either a locatum, a relatum, or a reference place in an inferred relationship, thus resulting in a total of six combinations. Finally, relative directions (either as already stored ones or inferred ones) with the same reference direction (i.e., having the same relatum, reference place, as well as the direction be- tween the reference place and the relatum) will be composed for con- sistency reasoning. The corresponding composition table of relative direction relations is similar to the one for cardinal direction relations as shown in Table 8.
inverse relationship inference Each one of the spatial rela- tions above, except qualitative distance relations, has a determined in- verse relation, shown in Table 12. The inverse relations for qualitative distance relations are not considered due to their context-dependency. For example, “the bike is near the garage” does not imply that “the garage is near the bike” (Landau and Jackendoff, 1993), and therefore
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 4.2 consistency reasoning in a place graph database 67 the inverse relation of near may or may not be near (either symmetric or antisymmetric). The table is used for deriving inverse relationships for the second case of forming cycles introduced in Section 4.2.1.
Table 12: Inverse relations of cardinal direction, relative direction, and topo- logical relations.
r inv(r) r inv(r) r inv(r)
d d N S F B m m S N R L e e E W B F i ct W E L R cB cv NE SW LF RB ct i NW SE RF LB cv cB SE NW LB RF o o SW NE RB LF
4.2.3 Single-Family Relational Consistency Maintenance By Cycles
The consistency checking approach below attempts to identify incon- sistent spatial relationships from each of the four spatial relationship families when cycles are formed. All other relationships are not yet covered and will be stored without consistency checking (e.g., non- binary ones). During a PGD transaction, any new input spatial rela- tionship between two places (i.e., a new triplet since only binary rela- tions are considered) will be compared to existing spatial relationship knowledge between the two places, either explicitly stored or inferred. First, two definitions are given:
Definition 3 (Existing Path) In a place graph database, an existing path (EP) between two places is a sequence of places and spatial relations connect- ing these two places.
EPs can be queried by a breadth first search (BFS) or depth first search (DFS). The number of EPs between any two places can be none, one, or more. Each EP together with a new input triplet connecting the two places forms a cycle and may result in an inconsistency state.
Definition 4 (Existing knowledge) An existing knowledge (EK) is a single spatial relation or a set of relations between two place derived from an EP [n1, r1, n2, r2, ... , nm], either through composition r1 ◦ r2 ◦ . . . rm−1 (if the number of edges in the EP, or the length of the EP, is greater than one) or not (if the length is one).
EPs will only be queried and composited for deriving EKs when there is a new triplet being added to a PGD, either through creation or update. Such a new triplet must be consistent with the derived EKs. For instance, assuming a PGD with four places A, B, C, and D and three spatial relations connecting AB, BC, and CD respectively, the EK along the path consisting the four places and the three relations
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 68 qualitativespatialreasoning
determines which relations are consistent between A and D. Accord- ingly, a new triple
Definition 5 (Consistent triplet) A new input triplet is consistent with the database if it is an element of every EK (for every EP) between the locatum and the relatum of the triplet.
Note that since the above relational composition rules are defined for each single spatial relationship family, EP and EK are computed for relations only from the family of the new input spatial relationship. Therefore, for example, given a new triplet with a cardinal direction relation, only EPs consisting of cardinal direction relations only will be composited for EK. If the triplet is consistent, it will be added to the PGD; otherwise an inconsistency will be flagged. The full consistency checking procedure is described in Algorithm 2.
Algorithm 2 Consistency checking algorithm. Input: Locatum, Relation, Relatum Output: Boolean 1: RelationFamily = getRelationFamily(Relation) 2: EPs = getEPs(Locatum, Relation, RelationFamily) 3: for EP in EPs do 4: EK = getEK(EP) 5: if Relation not in EK then 6: return False . Flag inconsistency 7: end if 8: end for 9: return True . Accept transaction
An illustrative example is shown in Figure 20. Two spatial relation- ship among three places already exist in a PGD:
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 4.2 consistency reasoning in a place graph database 69
Figure 20: An example of consistency checking of a new triplet based on existing path and the derived existing knowledge.
4.2.4 Extending the Framework to Cross-Family Reasoning
This section proposes an extension of the single family-based reason- ing framework to cross family situations. The major motivation is that some relational inconsistency cannot be detected considering each single family of spatial relationship family each time. For example, triplets of
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 70 qualitativespatialreasoning
Figure 21: Consistency matrix, with green indicating that the two spatial relation are regarded as consistent with each other.
tension is rather to allow observations and discussions being made when tested on real world datasets and compared to the previous sin- gle family approach, instead of proposing the consistency matrix as a formal model for cross family spatial reasoning. Hence, experiments for both single family reasoning as well as cross family reasoning will be conducted in Section 4.3. The proposed consistency matrix allows some flexibility. For exam- ple, when A is disjoint with B, A can at the same time be either near or far to B, and can be in any cardinal or relative direction relation to B. As another example, two features that overlap cannot be far from each other. Following logic, within a family of relations one would expect a strictly diagonal submatrix. However, allowing for language flexi- bility the consistency matrix is more relaxed, for example accepting NW being consistent with N or W. The remaining part of this section illustrates how the previously defined inference (composition) and consistency checking algorithm can be extended leveraging the consistency matrix. Here the purpose is to construct a composition table consisting of all the spatial rela- tions in the consistency matrix. Let f ∈ {ϑ, δ, κ, ρ}, where f is a single relation family (ϑ the topological relations family, δ the qualitative distance, κ the cardinal direction, and ρ the relative direction). Let also r ∈ U, where U is the universal relation comprising all families.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 4.2 consistency reasoning in a place graph database 71
Then fA,B stands for all relations from U that are consistent with rA,B and belong to the same family f. Also, ⊗ is used to distinguish from ◦. ◦ denotes composition between two relations as defined earlier; ⊗ is used between two relations rA,B and rB,C that can be from different families and result in MA,C, consisting of relations from U. In order to perform composition over multiple families, the following formula is used:
rA,B ⊗ rB,C = MA,C =
{ϑA,B ◦ ϑB,C} ∪ {δA,B ◦ δB,C} ∪ {κA,B ◦ κB,C} ∪
{ρA,B ◦ ρB,C}
If, for instance, rA,B = E and rB,C = d, then the row of E of the ma- trix provides the relations consistent with E: ϑE = {d, m}, δE = {n, m, f}, κE = {N, NE, E, SE, S}, and ρE = {F, R, B, L, LF, LB, RF, RB}. Similarly ϑd, δd, κd, and ρd can be obtained. Then MA,C is the union of all com- positions of individual families. When computing each fA,B ◦ fB,C in the formula above, it is pos- sible that either one of, or both fA,B and fB,C is the empty set {∅}; such situation occurs when a relation is not consistent with any rela- tion in another family. For example, if rA,B = overlap, then κo = {∅}, in other words no cardinal direction is consistent with overlap. Let S be a subset of a single family, then the following three cases can be distinguished:
• One of fA,B or fB,C is {∅}: let the other one be S then fA,B ◦ fB,C = {∅} ◦ S = S. For instance, δcn ◦ δn = {∅} ◦ {n} = {n}.
• Both fA,B and fB,C are {∅}, and rA,B and rB,C belong to different families: then the result is {∅}. For instance, ρi ◦ ρf = {∅} ◦ {∅} = {∅}.
• Both fA,B and fB,C are {∅} and rA,B and rB,C belong to the same family. Then three steps are needed. First, compute rA,C = rA,B ◦ rB,C. Then for each relation in rA,C obtain fr. For exam- ple, for δo ◦ δm, r = ϑ, while fr = δ. Therefore: ϑo ◦ ϑm = {d, m, ct, cv, o}, then δd = {n, m, f}, δm = {∅}, δcn = {∅}, δcv = {∅}, δo = {∅}, and finally δo ◦ δm = {n, m, f} ∪ {∅} ∪ {∅} ∪ {∅} ∪ {∅} = {n, m, f}.
Accordingly, a full composition table including 27 rows and 27 columns can be generated according to the algorithm. Composition of more than two relations is performed in the same way. For example:
i ⊗ n ⊗ N = {i ◦ d ◦ (d ∨ m)} ∪ {{∅} ◦ n ◦ (n ∨ F)} ∪
{{∅} ◦ {anyκ} ◦ (N ∨ NE ∨ NW)} ∪
{{∅} ◦ {anyρ} ◦ {anyρ}} =
{anyϑ} ∪ {n, f} ∪ {anyκ} ∪ {anyρ} = any
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 72 qualitativespatialreasoning
With the cross family composition table derived, the previously in- troduced consistency checking algorithm can be used for cross family consistency maintenance in exactly the same way as for single family scenarios.
4.3 implementation and experiments
The described reasoning approach has been implemented using the constructed campus PGD, as introduced in Chapter 3. When compar- ing to the Melbourne PGD, the campus PGD has richer connections among places and more cycles, and therefore is preferable in this chapter for experiments. A comparison of spatial relations from the four families and examples of their associated NL expressions are shown in Table 13, with some directional relations omitted due to their similarity in expressions to the shown ones. In order to be able to use these families, a mapping of spatial prepositions in a PGD to the formal relationships has to be applied. A classification schema as a look-up-table is leveraged, as also implemented in (Kim, Vasardani, and Winter, 2016). For example, the NL expression close to will be mapped to the qualitative distance relation near, and western will be mapped to the cardinal direction relation west. The numbers and pro- portions of spatial relations from each family are shown in Figure 22. Among the input spatial objects, Campus is a special one. It is rather a container of most other spatial objects in these place descriptions. Among all 731 triplets, 144 of them (approximately 19.7%) are related to Campus.
Figure 22: The numbers and proportions of spatial relations from each of the four spatial relation families from the campus place graph database.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 4.3 implementation and experiments 73
Table 13: Formal spatial relations and examples of their original natural lan- guage expressions from place descriptions.
Formal relations Natural language expressions
north north, N, northern, north from east east, E, eastern, east from west west, W, western, west from south south, S, southern, south from
near close to, near, nearby middle not far from, not very far, middle distance of far far from, away from, further
front in front of, before, in the front of, ahead of back behind, back side of, opposite of, other side of left left of, left to, left side of, left hand side of right right of, right to, right side of, right hand side of
disjoint outside of, separated from meet meet, touch, adjacent inside inside, center of, in, within equal same, equal, as, also known as contain surround, include, contain coveredBy covered by, at the end of cover internally connected to overlap overlap, intersect
4.3.1 Single-Family Reasoning
Figures 23, 24, 25, and 26 show the numbers of EPs found by place id for all places in the campus PGD for the four spatial relation families respectively. The number of EPs are shown considering EP length of 1, 2, and 3 (a new triplet with an EP of length n results in a cycle of length n + 1) in a cumulative manner. These EPs are queried us- ing the embedded traversal engine by Neo4j based on DFS. In each of the four figures, the left subfigure shows the number of EPs with- out considering inverse spatial relations (all spatial relations in an EP has the same direction in terms of sequential order of places, such as
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 74 qualitativespatialreasoning
for reasoning. The second way is not adopted since if a triplet con- tains incorrect spatial relationship knowledge and is not yet flagged as being inconsistent, the incorrect knowledge will propagate as the inferred spatial relationships that are based on this incorrect knowl- edge will be stored as well. Such inferred spatial relationships could potentially affect the following reasoning processes.
Figure 23: Number of existing paths of cardinal direction relationships queried by place id for all places in the campus place graph database, with (right) and without (left) considering inverse re- lationships.
Figure 24: Number of existing paths of relative direction relationships queried by place id for all places in the campus place graph database.
Pragmatically a maximum EP length has to be specified for consis- tency checking, and there is a trade-off problem when trying to define an appropriate maximum EP length. A larger value makes it possible to identify inconsistent knowledge that require more steps of com- position along longer EPs. However, a larger value also significantly increases the query time for searching EPs. A comparison of query time for individual places with different EP length settings is shown in Figure 27. The query time soars when the length exceeds 3. More- over, it is observed that after just a few steps of composition (typically 3), the inferred relations end up with any quickly anyway, reducing
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 4.3 implementation and experiments 75
Figure 25: Number of existing paths of qualitative distance relationships queried by place id for all places in the campus place graph database.
Figure 26: Number of existing paths of topological relationships queried by node id for all places in the campus place graph database. the urgency to test EPs with longer length. Therefore, the maximum EP length is pragmatically set to 3 for consistency checking in this thesis. In order to evaluate the consistency checking algorithm, all the triplets from the campus dataset are reconstructed to a new PGD, and during the triplet import stage, the algorithm is applied to identify inconsistent triplets. Each flagged triplet is then manually verified. Precision is used for evaluation of the algorithm, which is defined as the number of verified inconsistent triplets divided by the total number of identified inconsistent triplets. As results, no inconsistent triplets are found for the cardinal direc- tion, qualitative distance, and topological relationship families. This is not surprising, since the descriptions are from students that are already, to various extents, familiar with the environment of the Uni- versity of Melbourne’s Parkville campus, therefore the possibility of giving misleading descriptions is low. Four relative direction relations are identified as inconsistent. After verification, these relations are confirmed as being inconsistent to some existing relationship knowl-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 76 qualitativespatialreasoning
Figure 27: Query time (in seconds) for existing path with different lengths for individual places.
edge stored, and therefore the precision of the algorithm is 1 for the tested cycles. An example of inconsistent spatial relationships is shown below in Figure 28 with two descriptions:
“... You’re now in the Old Quad ... Pass through the Old Arts building and immediately look to your left—the tall building is the Babel building that, somewhat ironically, houses the languages and linguistics departments ...”
“... From the Old Quad, you can go through the Old Arts building, and then turn right and walk until you come to a building called the Babel building (a 1970s yellow brick monolith) ...”
Figure 28: The locations of the three places mentioned in the descriptions above, with a red arrow indicating the reference direction.
The two relationships between the Babel building and the Old Arts Building from the two different descriptions are denoted here as
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 4.3 implementation and experiments 77 when compared to the ground-truth configuration from the map in Figure 28. The algorithm successfully identified the two relationships as being inconsistent. The resulting precision is maximum as the original input descrip- tions are generally consistent. Therefore, in order to further test whether the consistency-checking algorithm is robust of rejecting all inconsistent triplets and accept all consistent ones, an additional set of 60 made-up consistent and (deliberately) inconsistent sets of triplets of all four categories of relationships is imported to the database con- structed in the previous experiment. Note that all the manually made triplets are imported in sequential order, and each set of triplets does not exceed the EP length limitation of 3. As a result, 60 out of 60 sets of triplets, both consistent and incon- sistent ones, are classified correctly, and thus the precision of the al- gorithm remains 1. The experiments show that the consistency check- ing algorithm is capable of distinguishing consistent and inconsistent triplets within the EP limitation for single family spatial relationships.
4.3.2 Cross-Family Reasoning
Up to now, each relationship is only checked for consistency within its own category. This section applies the reasoning framework in- troduced in Section 4.2.4 for consistency maintenance over multiple spatial relationship families. Figure 29 shows the numbers of EPs by place id for all places with all spatial relations considered, again for EP lengths of 1, 2, and 3. The numbers are generally more than 10 times larger than those for each single family as shown previously. The observed query time for EPs is also longer than the time for querying EPs over each single family. Again, with long EPs considered, any creation or update operation may become more costly in terms of computational time and memory consumption.
Figure 29: Number of existing paths of any spatial relationships queried by place id for all places in the campus place graph database, with considering inverse relationships (right) and without (left).
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 78 qualitativespatialreasoning
Based on the defined consistency matrix and the derived compo- sition table for all relations, it is anticipated that composing multi- family spatial relations easily results in the universal set of any.A comprehensive test is conducted by composing spatial relations in all possible combinations. The result shows that, when path length is set to 2, 28.9% of composition cases reach any. The numbers are 73.6%, 92.5%, 98.1% and 99.6% for path lengths of 3, 4, 5, and 6 respectively. Note that the percentage will never reach 100%, since there are cases such as north_of ⊗ north_of ⊗ north_of, which will result in north_of and other relations, but not any. This result supports the decision of choosing 3 as the maximum EP length for consistency checking con- sidering computational performance, knowing that composition of cross family relations quickly ends with any anyway over long paths. Next, the same two experiments from the last subsection for sin- gle family reasoning, i.e., consistency checking over triplets from the campus datasets when used for constructing a PGD, as well as the ad- ditional experiment using manually made triplets, are conducted. A total number of twelve relationships are flagged as inconsistent in the two experiments. Other than the inconsistent ones shown in the last subsection, the remaining ones can generally be classified into two categories, which may be regarded as false positives or not:
• Different interpretations of spatial relationships: For exam- ple, some descriptions produced
• Vagueness of the ‘boundary’ of places: For example
Therefore, while consistency rules can be set out from a logical per- spective, such rules do not always reflect pragmatics from a linguistic perspective, especially when context is involved.
4.4 chapter discussion and summary
This chapter presents a framework for spatial relationship consistency maintenance in a PGD. Four different families of spatial relationships are considered, building on the established field of QSR: cardinal di- rection, relative direction, qualitative distance, and topological rela- tions. The major tools are the defined relational composition tables as well as a path-based consistency-checking algorithm. The motivations of developing the framework is that that people may make mistakes when giving place descriptions, and malicious users could try to add
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 4.4 chapter discussion and summary 79 inconsistent knowledge. Since the source of knowledge to be repre- sented are human place descriptions, traditional frameworks based on formal logic may be inappropriate, and instead pragmatic choices were built into the design. The implemented consistency maintenance framework as the outcome does only flag an inconsistency as it is de- signed to; it has no mechanisms to decide whether the already stored or the new spatial relationship knowledge is the truer or closer to the representation of geographic reality. Composition rules, which are mostly novel, are defined for spatial relations from each single family, as well as for all spatial relation- ships together using a defined consistency matrix. Relational infer- ence for deriving possible spatial relationships between two places based on already stored knowledge then becomes possible, which is done by composing spatial relations along paths connecting the two places. For each triplet to be created or updated, the database tests consistency of the new relationship over any cycles of a limited length that will be formed by the added triplet. Thus whenever a transaction is being processed (i.e., creation, update, deletion), the place database will either end in a consistent state, or flag the new triplet for human inspection. For example, if the database contains a triplet
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 80 qualitativespatialreasoning
work must address the current weaknesses experienced in this thesis. The major weakness requires an even deeper consideration of context. While consistency rules can be set out from a logical perspective, such rules do not always reflect pragmatics from a linguistic perspective. The usage as well as interpretations of spatial relationships in NL is flexible and context-dependent. Going back to the previous example, “the bike is near the garage” does not imply equal acceptance of “the garage is near the bike”: there is a shift of context between the two statements. The same argument about context might also apply to two nodes referring to places of significantly different spatial extent, which immediately questions the straight applicability of the inverse relation for direction relationships. However, in the case of campus descriptions the spatial extents of features were comparable. Future work could also employ reasoning models based on other types of logic, such as fuzzy, probabilistic, or defeasible logic-based ones. This is expected to further increase the flexibility of the reasoning process through better modelling the semantics of NL spatial relationship ex- pressions.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] PLACEGEOREFERENCING 5
Place descriptions with references to places and spatial relationships provide spatial knowledge of places in their relative locations, and such knowledge has been captured in the proposed place graph database (PGD) model. Place descriptions often contain names refer- ring to fine-grained places, such as buildings or local point of in- terests (POIs), that are more difficult to disambiguate than names re- ferring to larger places. In addition, place descriptions may contain places not referred to by official names, or places not captured in of- ficial databases such as gazetteers. Places from these situations are more challenging to resolve and require different methods than pre- vious proposed ones for toponym resolution (TR). This chapter introduces a three-step georeferencing approach that is able to georeference all places from a PGDs constructed from col- lective descriptions, regardless of whether they are referred to by gazetteered names or not. The approach divides places in a PGD into three situations and georeference them accordingly: places re- ferred to by gazetteered names, gazetteered places not referred to by gazetteered names, and non-gazetteered places. While the first type of places is considered in TR studies, places from the other two sit- uations are typically ignored. The approach takes advantage of the structure of a PGD, specifically merged place references as well as spa- tial relationship knowledge, thus resulting in higher georeferencing precision and recall compared to conventional TR approaches. This chapter is based on content from previously published papers, including a novel clustering algorithm (Chen, Vasardani, and Winter, 2018a) and a three-step georeferencing approach (Chen, Vasardani, and Winter, 2018b). The contributions of the corresponding (first) au- thor include problem conceptualization, literature review, methodol- ogy, implementation, experiments, and paper writing.
5.1 introduction
Place descriptions typically provide a qualitative reference system for describing geographic locations, and consist essentially of references to places and their qualitative spatial relationships. In communica- tions, place references and spatial relations are used for conveying the locations of places, and can be used to, for example, provide navigational instructions or inform the location of events. Georefer- encing places in a PGD could increase the usefulness of the system through linking it to other geographic information systems (GISs), spa- tial databases, and spatial services for several application scenarios, such as geographic information retrieval (GIR), place searching, map- ping, and analysis.
81
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 82 placegeoreferencing
Previous approaches for georeferencing place from textual docu- ments are mostly based on TR of gazetteered, i.e., officially indexed, place names, and ignore references to places that are not. These ap- proaches also tend to focus on larger geographic features such as populated places (e.g., cities or towns) or natural geographic features (e.g., rivers or mountains). Everyday place descriptions, on the other hand, are often flexible, vernacular, and does not apply most exist- ing approaches from TR due to three reasons. First, place descrip- tions often contain place names of fine-grained features (e.g., names of streets, buildings and local POIs). Previous knowledge-based meth- ods designed for coarse places or natural geographic features may use disambiguation heuristics leveraging the prominence, population, or containment relationships based on external knowledge bases. Such heuristics quickly fail when dealing with the fine-grained places as they are often significantly more numerous and more similar to each other. Disambiguation approaches based on machine-learning tech- niques are difficult to be applied for fine-grained places due to the lack of good-quality training data, as well as the challenge of locat- ing previously-unseen place names. Second, place descriptions often contain place referents that cannot be found in a gazetteer, such as synonyms or place type references (e.g., ‘the large square’). Gazetteer matching quickly fails when facing such references. Last, some places may not be captured by gazetteers, such as references to places that are from too fine-grained environments to be captured by gazetteers (e.g., ‘the dean’s office’), and references to vague places or vernacu- lars that exist only in limited contexts (e.g., ‘the BBQ area near the tree in front of our department’). In communications, such places are typ- ically located by providing spatial relations to some landmarks with locations better-known. These three reasons result in most of these existing approaches from the TR community not directly applicable for the task of this chapter: georeferencing all places in a PGD, regardless of whether they have been referred to by gazetteered names or not. This chap- ter illustrates a three-step approach for georeferencing all places in a PGD. The first step attempts to georeference places that have been referred to by gazetteered names. For this purpose, a novel clustering- based disambiguation algorithm is developed, which is superior com- pared to competitive clustering algorithms as it does not require man- ual input parameters and is flexible for disambiguating places from place descriptions with different conversational contexts. The second step leverages spatial relation search space models on the stored spa- tial relations between the remaining places (i.e., places not referred to by any gazetteered names) and places resolved in the first step to derive approximate location representations for these remaining places. Finally, a weighted multi-value similarity measuring approach is presented for matching places from the second step to gazetteer en- tries, based on their derived approximate location representations as spatial constrains. Even if the matching fails, i.e., the place is non- gazetteered, the derived approximate location representation from
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.2 a three-step georeferencing approach 83 the second step can still be used to visualize the location of this place on a map. The remainder of this chapter is structured as follows. The three- step approach, including the novel clustering algorithm, spatial re- lation search spaces and approximate location regions, as well as gazetteer similarity matching is explained in Section 5.2. The imple- mentation, experiment results, and evaluation for each of the steps in terms of precision and distance error are provided in Section 5.3. Specifically, the clustering algorithm is evaluated against competing algorithms for the first step, different search space models are com- pared for the second step, and different parameters are tested for gazetteer matching for the third step. A comparison of the georefer- encing approach using the proposed PGD model as well as the ba- sic place graph (PG) model has been provided as well. Section 5.4 presents a discussion of the obtained results and major observations as chapter summary.
5.2 athree-step georeferencing approach
This section first clarifies three core subtasks and provides an overview of the workflow. Then, detailed method for each of the core subtasks is explained in the following three subsections.
5.2.1 Overview
Each place node in a PG has a unique identifier and at least one, but is potentially linked to multiple place references, as between places (as conceptualized in the real world) and place references are n:m re- lationships. Figure 30 shows an example representation of six places represented by nodes (labeled a, b, c, d, e, f ) linked by seven spatial relationships represented by labeled edges. Note that this example is only for illustration of the problem, instead of an actual PG. A list of place references from the original place descriptions for each node is shown in the solid-line rectangles. Each dashed-line rectangle shows the ground-truth gazetteered name(s) for these places (‘-’ for non- gazetteered places). Note the difference between a non-gazetteered reference and a non- gazetteered place: a non-gazetteered reference may either be a synonym referring to a gazetteered place, or a reference referring to a non- gazetteered place, while a non-gazetteered place does not have any corresponding gazetteer entries. Thus, three situations can be distin- guished for places in a PG (taking Figure 30 as example):
1. A gazetteered place with at least one gazetteered reference and possibly other non-gazetteered references (nodes a, c, d).
2. A gazetteered place with no gazetteered references (nodes b, e).
3. A non-gazetteered place (node f ).
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 84 placegeoreferencing
Figure 30: A sample representation of places linked by spatial relations. Each place is associated with one or more place references cap- tured in a place graph.
According to the task of this chapter, places from these three situ- ations must all be resolved. The solution provided below is divided into three main phases, as shown in Figure 31. The first phase aims to identify and disambiguate places from Situation 1 and use them as anchors. In the second phase, spatial relations to the anchors are leveraged for approximate localization for places from Situations 2 and 3. Finally, gazetteer matching based on the derived approximate location representations from the second step is conducted to refine the georeferenced location of places from Situation 2. Accordingly, the georeferencing result can either be an entry from a gazetteer with a geometrical representation (typically a point) for places in Situation 1 and 2, or a density surface for places from Situation 3.
Figure 31: The workflow of this chapter, with the first three phases corre- sponding to the three major subtasks of the approach.
5.2.2 Step One: A Clustering Algorithm for Disambiguating fine-grained place names
In the first phase, all place references in an input PGD are looked up using a gazetteer. If a place has at least one associated place ref- erence that can be found in the gazetteer, it is regarded as an an-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.2 a three-step georeferencing approach 85
chor place (i.e., Situation 1). Given i gazetteered places {p1, p2, ... , pi}, each place has a set of (one or more) corresponding ambiguous en- 1 2 j tries in the gazetteer {pi , pi , ... , pi} that the captured place references of this place can be matched to. The task of this step is to assign each anchor place with one entry from its ambiguous gazetteer en- 2 tries (e.g., pi to pi ) through disambiguation. For example, the place (p1) referred to by only one reference name Melbourne refer to multi- 1 ple places on earth, such as Melbourne, Victoria, Australia (pi ), or Mel- 2 bourne, Florida, US (pi ). The task of disambiguation is to determine 1 2 whether p1 is corresponding to pi or pi . For this step, map-based disambiguation approaches (typically clustering) are preferred as they are least demanding in terms of knowledge required. There are several clustering algorithms used in the literature for place name disambiguation (e.g., Buscaldi and Rosso, 2008a; Habib and Keulen, 2012; Moncla et al., 2014). How- ever, these algorithms may not be suitable for the task of this chap- ter. Some of them are defined for large geographic features and may not perform equally well on fine-grained places. Some algorithms are parameter-sensitive, and require manual input, and thus substan- tial a-priori knowledge of the data. Place descriptions are potentially from various conversational contexts and therefore no a-priori knowl- edge can be assumed. Therefore, in this section a novel density-based clustering algorithm DensityK is proposed. The algorithm is designed to be robust, parameter-independent and -insensitive, and it does not require manual input parameter values. The algorithm consists of three steps that are explained below. It will be evaluated against com- petitive algorithms in Section 5.3.2.
5.2.2.1 Computing Point-Wise Distance Matrix In the first step, DensityK computes all point-wise distances of an in- put point cloud (locations of all ambiguous entries of places from an input dataset), and the time complexity is O(n2) (n is the number of input points). The time complexity can be reduced to O((n2 − n)/2) with a distance dictionary to avoid re-computation (but needs O(n2) memory). Point-wise distance matrix has been used by several clus- tering algorithms in the data mining community, such as DBSCAN (Ester, M., Kriegel, H. P., Sander, J., & Xu, 1996).
5.2.2.2 Deriving Cluster Distance In the second step, DensityK analyzes the computed point-wise dis- tances, and derives a cluster distance automatically. The cluster dis- tance will be used in the next step for generating clusters. First, a DensityK function is computed given the point-wise dis- tances in the first step, as shown in Function 4. K(d) represents the average point density for points within a given distance inter- val (d − ∆d, d] for all points in an annular region. The reason to apply annular search region for computing point density instead of circular region (i.e., ∆d = d) will be explained below in this section. In Func- tion 4, the expression count(p ∈ region(pi, (d − ∆d, d])) represents
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 86 placegeoreferencing
the number of points that are at a distance between d − ∆d and d (including d) from point pi. If there is no point within all the search regions for all points for a distance interval (dj − ∆d, dj], skip to the next interval ((dj, dj + ∆d]). Thus, K(d) is aways positive. The denom- inator of the left side of the function is the area of the annular region. ∆d is for discretizing the function and is set to 100m in this chapter. The resulting cluster distance threshold will be the integer multiple of ∆d. A demonstration will be given below in this section that the clustering result is little sensitive to the value of ∆d.
1 1 n K(d) = × count(p ∈ region(p , (d − ∆d, d])) πd2 − π(d − ∆d)2 n i i=1 X (4)
The approach is inspired by Ripley’s K function (Ripley, 1976) which was originally designed to assess the degree of departure of a point set from complete spatial randomness, ranging from spatial ho- mogeneity to a clustered pattern. Ripley’s K function cannot be used to derive clusters nor cluster distances, yet the idea of detecting point density accordingly to distance threshold meets the interest of this chapter. The goal of this georeferencing step is to derive a cluster dis- tance threshold which leads to clusters with significantly large point densities. Such clusters are likely to be corresponding to the actual spatial contexts of the original descriptions, i.e., spatial extents where the descriptions are embedded. DensityK is a novel algorithm with a different purpose than Ripley’s K function, but Ripley’s K function can be regarded as a cumulative version of the DensityK function. If the point-wise distances from the last step are sorted, the time com- plexity of computing DensityK function is O(n) as it makes at most n comparisons regarding different values of d. The function is used to identify values of d with significantly large point densities. Two illustrative examples are given in Figure 32 (a) and (b) with different input data. For each of the two sample func- tions, K(d) starts at a non-zero value for the first d: 100m (the value of ∆d), which means there are some points that are within 100m from other points in the input point cloud. As d grows, the value of K(d) continues to decrease. For different input data, it is also possible that K(d) starts from a low value, and then increases until a maximum value is reached, after which it starts to decrease again. Next, the mean µ and standard deviation σ of all K(d) values (a finite set since the function is discretized by ∆d) are calculated. Then, the 2σ rule is applied, and the minimum value of d is selected as the
cluster distance, that is d > d0, d0 = argmaxd K(d) and K(d) = µ + 2σ. The derived cluster distances are also shown in Figure 32 (a), (b). Intuitively, the cluster distance is the value of d at the ‘valley’ of a DensityK - a visually identifiable (at least roughly) x-value where the decrease pace of K(d) value dramatically changes, leading to values close to zero. It is found that the resulting cluster distances always sit somewhere at the ‘valley’ of the functions (in terms of K(d) values)
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.2 a three-step georeferencing approach 87
Figure 32: Two example DensityK functions from different input data with cluster distance highlighted (a, b), and comparisons of DensityK functions generated based on annular and circular search regions for the same data as in (a) and (b) respectively (c, d). for different input data, and the derived clusters afterwards match quite well to the actual spatial contexts. A comparison of annular and circular (replacing all ∆d by d in Function 4) search regions is shown in Figure 32 (c) and (d), with the same input data as in (a) and (b) respectively. When tested on sample data, it is found that when applying annular regions, the de- rived clusters are always more constrained (as the computed cluster distances are smaller) and closer to the actual spatial contexts than those derived from circular regions. Such more constrained clusters are preferred as they are more likely to exclude ambiguous entries. It is found that they lead to higher disambiguation precision on the tested data as well. This phenomenon is most likely because when applying annular regions, the DensityK functions are more sensitive to the change of local density. In comparison, applying circular re- gions results in smoother density functions and possibly much larger cluster distances derived. DensityK function is little sensitive to the only parameter value of ∆d, which does need to be determined manually for different in- put data. As shown in Figure 33, the DensityK function plots gen- erated for the same input data with three different ∆d values 100, 250, and 500m are similar, and the cluster distances derived are ex- actly the same. ∆d should be set to a constant, small number (e.g., the values in Figure 33) for all input data, just for the purpose of dis- cretization. Such a small number works well for various input data, even those with large cluster distances. Note that there is no single-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 88 placegeoreferencing
optimal cluster distance for disambiguation. For example, different cluster distances from 2500m and 3500m may lead to the same dis- ambiguation result for a given input; however, a cluster distance with value 25000m for the same input may increase or reduce the disam- biguation precision, depending on the distances between the actual locations of the place names.
Figure 33: DensityK function generated with four different ∆d intervals for the same input point cloud: 100, 250, 500 (meters).
Algorithm 3 explains the whole procedure of this step, with sorted point-wise distances from the last step as input. The first part of the algorithm computes K(d) for different d values, and stores tuples of (d, K(d)) in the list variable KFunction. Then, the cluster distance is derived given KFunction.
Algorithm 3 Computing cluster distance threshold. Input: PointWiseDistances: a sorted list of distance floats in meters Output: ClusterDistance: a float in meters 1: KFunction := an empty list of 2-element tuples 2: MaxDistance := maxValue(PointWiseDistances) 3: NumberOfDistances := length(PointWiseDistances) 4: for d in iterate(0, MaxDistance, ∆d) do . loop of (min, max, interval) 5: PointCountInRadius := 0 6: for distance in PointWiseDistances do 7: if distance 6 d then 8: PointCountInRadius += 1 9: end if 10: end for 11: if PointCountInRadius > 0 then 12: Area := π(d2 − (d − ∆d)2) PointCountInRadius 13: Density := Area × NumberOfDistances 14: KFunction ← (d, Density) . Function 4 15: end if 16: end for 17: 18: Densities := getDensities(KFunction) 19: Mean, StandardDeviation := getMeanAndStd(Densities) 20: ThresholdDensity := Mean + 2 × StandardDeviation 21: ClusterDistance := getCorrespondingDistance(ThresholdDensity, KFunction) 22: return ClusterDistance
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.2 a three-step georeferencing approach 89
5.2.2.3 Deriving Clusters and Disambiguation The procedure of deriving clusters is similar to DBSCAN. Points that are within the cluster distance threshold are merged into the same cluster. The last step is to assign each place name with a location for disambiguation. To do so, the derived clusters are ranked by their contained number of points in descending order. Then, for each place name, choose the entry that first appears in one of the cluster accord- ing to the ranking, and the first cluster an entry appears is called a top-cluster for this place name. For example, if an entry of a place name appears in the cluster with the largest number of points, the entry will be selected for disambiguation. If no corresponding entry of the place name is found in the first cluster, then the next cluster is chosen, until one entry is found. Thus, the worst case time complexity of this step is O(nm) (m is the number of clusters derived). In prac- tice, as most places names are expected to be located within the first cluster, the time complexity is close to O(n). The reason that multiple clusters derived instead of only the first cluster is considered is be- cause it is possible that the input place names are from multiple spa- tial foci, i.e., the locations of some of the named places are relatively far away. In such cases, these isolated place names will be missed by the first cluster thus cannot be disambiguated correctly. The complete disambiguation procedure of this step is given in Algorithm 4.
Algorithm 4 Disambiguation using the derived clusters.
Input: Clusters, PlaceNamesAndEntries as an list of 2-element tuples {(pi, entryij), ...} Output: DisambiguatedPlaceNames 1: DisambiguatedPlaceNames := ∅ 2: RankedClusters := rankDescendent(Clusters) 3: for Place in getPlaces(PlaceNamesAndEntries) do 4: for Cluster in RankedClusters do 5: for Entry in getCorrespondingEntries(Entry, PlaceNamesAndEntries) do 6: if Entry in Cluster then 7: DisambiguatedPlaceNames ← (Place, Entry) 8: Goto 3 9: end if 10: end for 11: end for 12: end for 13: return DisambiguatedPlaceNames
Figure 34 illustrates an example of disambiguating anchor places from the example shown in Figure 30 through clustering. The dashed circle indicates approximately the spatial context (a derived cluster is a set of entries instead of an actual circle), where the disambiguated three entries shown on the map are within. Lastly, an additional process for further disambiguation is intro- duced, as it is possible that a cluster may contain more than one entry for an anchor place. In this case, these places are temporarily removed from the anchor place list, and will be geo-referenced together with the remaining places in the next phase, where spatial relationships will be used for further disambiguation.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 90 placegeoreferencing
Figure 34: Disambiguated anchor places a, c, d from the example shown in Figure 30, which are within the cluster identified as spatial con- text. (source: Google Map, 2015).
5.2.3 Step Two: Deriving Approximate Location Region Representations
Spatial relationships in place descriptions provide useful information for place localization. In a triplet, the locatum is located through pro- viding a spatial relation to the relatum, which is typically a landmark whose location is better known, and thus, in relation to which the location of the locatum can be specified. Yet such a relationship does not explicitly provide any locative information of the locatum, and thus, cannot be immediately interpreted by a computer without be- ing modelled. Consequently, although it is usually easy for a human recipient of a description to disambiguate place references or approx- imately locate places by spatial relationships, computationally mod- elling natural language (NL) spatial relationships to allow automatic interpretation remains a significant and open challenge. This section first introduces search spaces for spatial relations, ei- ther as formally defined ones or as contextualized probabilistic-based ones trained from data. Approximate location regions will then be in- troduced for integrating different search spaces in order to approx- imately locate the remaining places from the previous phase using density surface. Density surface-based representations have been fre- quently applied to places with indeterminate locations or boundaries (e.g., Gao et al., 2017a; Jones et al., 2008a), which are typically derived from collective location data (e.g., geotagged images or texts from so- cial media). In comparison, this chapter shows how the locations of places can be derived as density surfaces based on collective spatial relationship knowledge with search space models.
5.2.3.1 Formal Search Space Models The semantics of binary qualitative spatial relations from four fami- lies are considered, and these spatial relations stored in a PGD have been formalized already, as introduced in Section 4.3. A search space is defined for each relation from the four families to represent the constrained location of a locatum that satisfies the spatial relation to
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.2 a three-step georeferencing approach 91 an already geo-referenced anchor place (relatum). The search spaces below are defined for anchor place geometries of points, polylines, or polygons, although in most gazetteers places are represented by points. cardinal direction relations The search spaces for cardi- nal direction relations are defined based on Frank’s half-plane mod- els for point type relatum (Frank, 1992), as shown in Figure 35 (a, b, c). The model can be extended to support polygon-based relata (Fig- ure 35 (d)). However, in this chapter the centroid of any polygon is used to derive the half-planes. The reason is that a cardinal direction preposition might express an internal relation (Liu et al., 2005), e.g., ‘in the north (in the northern part) of the city’, which will be misinter- preted using the polygon-based model.
Figure 35: Search spaces for cardinal direction relations based on the cen- troid of the relatum (a, b, c), and an alternative model for non- point relata (d) that is not applied in this chapter. qualitative distance relations The search spaces for qual- itative distance relations are defined by buffer regions as shown in Figure 36 (a, b, c) for different relatum geometry types, similar to the ones proposed by Liu et al. (2009b) as shown in Figure 36 (d). Buffer regions are a generally accepted model for quantifying qualitative distances in applications such as in local-search applications or GIR engines. The buffer distances, which are highly context-dependent, are defined here empirically, and then adapted to the semantic con- text considering the size of the relata as well as to the size of the spatial context (which will be introduced in Section 5.2.3.3), as shown in Equation 5. d stands for the buffer distance, α is a constant, and β, γ are two coefficients that make d positively correlated with the area of the relatum, as well as the area of the spatial context.
Figure 36: Search spaces for the qualitative distance relations in this chapter (a, b, c), and a comparison to the model by Liu et al. (d).
d = α + β ∗ getArea(relatum) + γ ∗ getArea(spatialContext) (5)
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 92 placegeoreferencing
relative direction relations Search spaces for relative di- rection relations are defined based on orientation reference frames used by people, and can be either deictic, intrinsic or extrinsic (Retz- Schmidt, 1988). If the reference direction used is known (i.e., available in an input PGD), the search space of each relative direction relation is defined as shown in Figure 37. The arrow in the figure indicates the reference direction as front. However, current NL parsers are unable to infer the reference frames of prepositions automatically and accurately from place de- scriptions. If spatial reference frame knowledge is unavailable, the search spaces is set to be the same as near, as a fall-back approach.
Figure 37: Search spaces for relative directions given a reference direction.
topological relations If the relatum is polygon-based, the search spaces for different topological relations are defined as shown in Figure 38, otherwise no search spaces will be enforced. These search spaces are used as initial filters. In the later best-matching stage, topological relations will be further validated through geom- etry computation for excluding unsatisfactory gazetteer entries.
Figure 38: Search spaces for covered by, equal, inside (a), disjoint, meet (b), and the other three topological relations overlap, cover, contain (c).
5.2.3.2 Contextualized Probabilistic Search Space Models As refinements to the formal search space models, contextualized probabilistic search space models are introduced. Search spaces in this section are derived from training data and are contextualized by four factors, instead of formally defined. The values of the factors are only dependent on the input information introduced previously, and thus are automatically obtainable.
granularity of the relatum The semantics of a triplet’s re- latum can affect the interpretation of the triplet’s spatial relation. For instance, ‘near a restaurant’, ‘near an airport’, and ‘near Melbourne’ should be interpreted differently for defining search spaces. This fac- tor has also been used in the formal models for qualitative distance relations introduced above. Here the semantic types of relata are grouped into five categories based on spatial granularity, inspired by
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.2 a three-step georeferencing approach 93 the classification proposed by Richter et al. (2012): finer than building, building-, street-, district-, or city-level and beyond city. The underlying assumption is that places from the same spatial granularity level gen- erally have similar search spaces for spatial relations. granularity of the locatum Similarly, the semantics of the locatum can also affect search spaces. For example, ‘the building is near the CBD’ and ‘the suburb is near the CBD’ differ in their search spaces for the locatum. The theory of contrast sets by Winter and Freksa (2012) offers an explanation for this. The contrast sets of the two locata are other buildings and other suburbs in the current con- versational contexts, respectively. Therefore, the search spaces for sub- urbs will be larger than for buildings. The same five categories of place granularity are used for this factor. prominence of the relatum Landmarks are cognitively salient spatial objects in terms of prominence and distinctiveness (Richter and Winter, 2014) and are often used to locate other, less prominent places. Thus, landmarks should ideally have larger search spaces considering their influences compared to less prominent places from the same granularity. The degree of prominence of a rela- tum can be measured by the frequency of references to this place in all collected place descriptions. Also, prominence can be discretised using a two-valued logic: prominent and not prominent. granularity of the spatial discourse This factor is similar to the scale effect identified by Yao and Thill (Yao and Thill, 2005), introduced here as the granularity of the spatial discourse. For example, the relation near in the description ‘near Eiffel Tower’ can be interpreted differently in different discourses. For example, a place description could completely be located to a limited area near the tower, or could cover the whole city of Paris. For a triplet, the granularity of the spatial discourse can be obtained by first collecting all places from the same description (which has been indexed) and selecting the coarsest granularity category (from the named five) among these places, or, if all places are of the same category, one level up. Consider the example ‘Richmond is near the CBD’ (both are from district-level). In this case, it makes sense to limit the spatial discourse to city-level, since neither a suburb nor a city’s centre can be larger than the city that contains both.
A combination of contextual factor values is called a contextual cri- teria set (CCS), e.g., {granularity of the relatum: building, granularity of the locatum: building, prominence of the relatum: prominent, granular- ity of the discourse: district}. For each of the four contextual factors, an additional value undetermined is defined in case a value cannot be determined. A spatial relation will have one search space derived for each possible CCS. Using this method, search spaces for relations even not included in the formal ones discussed above (e.g., at) can be derived.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 94 placegeoreferencing
For a given triplet with relatum as an anchor place, the following approaches are used to associate it to one of the defined CCSs. The granularity of the relatum will be determined by mapping the stored place type of the relatum in the gazetteer (which is typically from a taxonomy) to one of the six granularity categories by a dictionary. The granularity of the locatum will be determined similarly, through identifying place type keywords from the stored place references of the locatum. For example, if keywords such as ‘building’, ‘park’, or ‘city’ occur, the locatum will be assigned with a granularity level ac- cordingly. If dictionary matching fails, e.g., the granularity of the ref- erence ‘the place’ cannot be determined, the value of locatum gran- ularity will be undetermined. Since the granularities of all places (as locata or relata embedded in triplets) can be determined, the granu- larity of the spatial discourse can be derived as well based on the rule defined above. For the prominence of the relatum, node in-degree as the number of references made to particular places are used, similar to the approach proposed by Kim, Vasardani, and Winter (2017a). To translate absolute measures of prominence (in-degrees) into relative measures of significance (prominent and not prominent), the median in-degree value is used as threshold. In the remaining part of this section, three different models for representing search spaces are proposed.
density surface model The first model is based on kernel den- sity estimation (KDE), a non-parametric method to estimate the proba- bility density function of some observation data. It provides a tool to visually represent vague concepts, and each region on the generated density surface represents the relative likelihood of a new observation within it. Figure 39 (a) shows an example search space generated through KDE for near given a CCS: {granularity of the relatum: building, granu- larity of the locatum: building, prominence of the relatum: not promi- nent, granularity of the discourse: street}. Assuming a set of training triplets that satisfy the CCS, for each triplet, the location of the loca- tum is regarded as a position vector relatively to the location of the relatum and mapped on a 2D plane. The result is a point cloud as the input of KDE. The generated density surface provides an intuitive representation of the search space of near of the CCS.
regression model The regression model aims at smoothing the density surface generated through KDE and avoid over-fitting. For this purpose a Gaussian process regressor (GPR) is used. A Gaussian Pro- cess is a generic supervised learning method designed to solve regres- sion and probabilistic classification problems. The prediction interpo- lates the observations and is Gaussian probabilistic, and thus allows for deriving meaningful approximate location regions. Another rea- son for applying GPR is that, individuals use and understand spatial relation phrases differently, and thus results in multi-component dis- tributions aggregated for search spaces. Figure 39 (b) shows the result after regression using the same data.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.2 a three-step georeferencing approach 95
Figure 39: Example of the density surface generated by KDE (a), the density surface generated by regression (b), and the hexagon representa- tion generated by tessellation (c, d) for near trained for a spec- ified contextual criteria set, based on relative locatum locations (distance [m]). tessellation-based model As the second model, a hexagonal partition of the space is defined. Examples (with different cell sizes) using the same training data are shown in Figure 39 (c, d). The model generalizes certain details and has a reduced computational complex- ity compared to the KDE model when used in the next phase. Choos- ing different cell sizes can affect the generated search spaces, as can be seen from the two sub-figures. Generally, a smaller cell size will more likely result in more dynamics (less smoothing).
5.2.3.3 Approximate Location Region
An approximate location region (ALR) is a derived region that repre- sents the approximate location of a place based on all known spatial relationships to some anchors, and is computed by integrating all the search spaces of these relationships, as well as the spatial context by intersection. The spatial context is defined, in this chapter, as a buffer region based on the minimum bounding box of the locations of all the anchor places, with a buffer distance equal to the cluster distance. A default spatial context is acting as a fall-back approach to locate places if they do not have any available spatial relationship knowl- edge to anchor places. An example of a spatial context is illustrated by the dashed circular region shown in Figure 34.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 96 placegeoreferencing
For the sample places shown in Figure 30, integrating different search spaces as formal models for deriving ALRs for nodes b and e is illustrated in Figure 40 (a). Place b from the example has no gazetteered references; however, three relationships are available, i.e., east of a, south of c, and near c. Knowing that a and c are already georeferenced in the previous phase, the location of b can be con- strained by the shaded region representing the ALR where b is most likely within. Thus, gazetteer entries outside the region will not be considered for matching in the best-matching phase (the third step). Integrating search spaces as probabilistic density surfaces based on KDE-, tessellation- and regression-based models, leads to slightly dif- ferent ALRs. Given n search spaces generated by the KDE or regression models, Equation 6 presents a product operation for integration. In the equation, s(x, y) stands for the value of a search space at location (x, y), and p(x, y) stands for the value of the derived ALR at location (x, y). The value of p(x, y) represents the relative likelihood of a place to occur at that location. For the hexagon tessellation model, s(x, y) is instead computed by the number of observation points within the cell divided by the total number of points in the input point cloud. Figure 40 (b) illustrates the integration process for two search spaces generated by KDE (the blue and green contour lines) into an ALR den- sity surface. In order to use the new search spaces in the later georef- erencing process, the values in such an ALR are normalized between 0 and 1. If a crisp boundary is required (e.g., for visualization), a threshold value for membership can be selected.
n p(x, y) = si(x, y) (6) i=1 Y
Figure 40: An example of deriving the ALR (the shaded region) for Place b through integrating three search spaces (Source: Google Maps, 2015) (a) and deriving an ALR by integrating two probabilistic search spaces into a new density surface (b) (distance [m]).
5.2.4 Step Three: Gazetteer Best Matching
In the third phase, ALRs are used for attempting gazetteer entry matching. This is done by first collecting all gazetteer entries within the ALR of each of the remaining place, and then choosing the one
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.2 a three-step georeferencing approach 97
that is most likely to be the actual entry the place reference addresses. Three measurements (each between 0 and 1) are considered for best matching.
reference string similarity Reference string similarity mea- sures how well a place name from a candidate gazetteer entry matches a place reference in string. Existing algorithms are many, and a comprehensive comparison for toponym matching has been pro- vided (Recchia and Louwerse, 2013). The selection of the algorithm is dataset dependent, and in this chapter the Damerau-Levenshtein al- gorithm is used (Damerau, 1964; Levenshtein, 1966). It is a commonly used algorithm for matching tasks such as gazetteer conflation or POI matching in the literature. It is expected to perform well on the test dataset since many place references from the dataset are short, incom- plete, and vernacular.
semantic similarity Semantic similarity measurements have been extensively studied in communities such as information re- trieval. In this chapter the Jiang-Conrath distance (Jiang and Con- rath, 1997) over WordNet synsets as lexicons is used for measuring semantic similarity word-wise, e.g., ‘woods’ and ‘forest’, ‘department’ and ‘section’. It is a common algorithm and similar implementations for other tasks already exist (e.g., Ballatore, Bertolotto, and Wilson, 2013). Abbreviations (e.g., ‘bldg’ vs. ‘building’) are considered as hav- ing semantic similarity that equals to 1. Additionally, place type key- words associated with gazetteer entries are considered as well to as- sist matching, if available. Taking the gazetteer of OpenStreetMap (OSM)1 for example, tagging information is stored with most entries, e.g., {name: Peter Hall Building; type: building; organization: unimelb; department: Mathematics}. The highest word-wise semantic similarity value will be returned.
spatial relation satisfaction Spatial relation satisfaction is for measuring how well a gazetteer entry at a certain location satisfies the given spatial relationships. For formal search space models, this is computed considering orientation, distance, and topology. For exam- ple, if two entries obtained for the place reference the large square for node b in the sample input PGD have the same name, they can only be ranked by their closeness to the anchor place St Paul’s Cathedral given the spatial relationship near. Methods for computation are shown in Figure 41. The shaded re- gions indicate search spaces. Nearness satisfaction is measured by the distance between the locations of the entry and the relatum, and must be between 0 (furthest) and 1 (closest). Orientation satisfaction is mea- sured by the angle between the displacement vector starting from the relatum to the entry location, as well as the direction specified by the relation (1 for 0◦ and 0 for 90◦). Topology satisfaction is measured by computing the topological relation between the two places, and can
1 https://www.openstreetmap.org/
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 98 placegeoreferencing
be either 0 (not satisfied) or 1 (satisfied). If an entry does not satisfy a given topological relation constrain, the entry will be excluded imme- diately. Topological relation computation can be implemented using existing libraries with models such as DE-9IM (Strobl, 2017).
Figure 41: Illustration for spatial relation satisfaction for near with relatum in the middle (a), north of with relatum at the bottom (b), and overlap with relatum in the middle (c).
For contextualized probabilistic-based search spaces, an ALR repre- sents the likelihood of a locatum being at different locations. Thus, the value for spatial relation satisfaction for a given candidate entry is simply the value of the ALR at the location of the entry. The values have been normalized and, thus, must be between 0 and 1.
overall scoring For each candidate gazetteer entry, the over- all score is calculated by Equation. 7 in a weighted multi-attribute manner. Different weights will be tested in the implementation stage. Table 14 shows an example of calculating the overall scores for three candidate entries for node b. For each of the two place references Fed Sq. and large square stored with node b, values of the three mea- surements for the three candidate entries (Ian Potter Centre, Federation Square, and Kirra Galleries) are calculated. After overall scoring, the highlighted cell in the last column of the table, i.e., Federation Square with the highest score 0.7, will be used for geo-referencing node b.
OverallScore = W1 ∗ StringSim + W2 ∗ SemanticSim + W3 ∗ SpatialSat (7)
Place Place reference Candidate entry Overall score
node b Fed Sq. Ian Potter Centre 0.37 Federation Square 0.70 Kirra Galleries 0.22 large square Ian Potter Centre 0.43 Federation Square 0.63 Kirra Galleries 0.27
Table 14: Example of best-matching for node b based on computed overall scores.
At the end of this phase, a score threshold is necessary to decide whether the matching process was able to find a gazetteer entry. Dif- ferent threshold values will be tested in the implementation stage. A non-gazetteered place, such as node f from the sample graph, will
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.3 implementation and experiments 99
then be geo-referenced only by its ALR. With such a representation, the location of the place can further be described using anchoring the- ory (Galton and Hood, 2005). Thus, the place can be regarded as an- chored to a location just by stating what is known with certainty and leaving the rest for further reasoning. Here, the place can be described as anchored in its derived ALR.
5.3 implementation and experiments
The described approach has been implemented using the constructed two test PGDs introduced in Chapter 3. This section is divided into three parts. In the first part, the developed clustering algorithm Den- sityK for the first step is evaluated against competitive clustering al- gorithms already exist, in order to demonstrate its superiority over the task of disambiguating place names from place descriptions. The experiment results for each of the three steps are presented and dis- cussed in the second part of the section. In the third part, the georefer- encing results using the proposed (extended) PGD model is evaluated against using the basic PG model for comparison.
5.3.1 Preprocessing
Three gazetteers were used in conjunction for retrieving entries for places in the test datasets, aiming for completeness: OSM Nominatim geocoder 2, GoogleV3 geocoder 3, and GeoNames 4. For instance, the name St Margaret’s School has a total of 11 corresponding entries from the three gazetteers. The retrieved entries from the three sources were then synthesized, and duplicated entries referring to the same places were removed. The numbers of ambiguous gazetteer entries retrieved are shown in Figure 42, representing the ambiguities of these place names.
Figure 42: Numbers of ambiguous gazetteer entries of places names from the two datasets, campus (left) and Melbourne (right).
Next, all places are manually linked to the corresponding gazetteer entries to create the ground-truth data for evaluation. Note that this
2 https://nominatim.openstreetmap.org/ 3 https://developers.google.com/maps/documentation/geocoding/intro 4 http://www.geonames.org/
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 100 place georeferencing
is not only for places from Situation 1 (anchor places), but also for places from Situation 2 (gazetteered places not referred to by gazetteered names). Places from different situations are also labeled, as they will be evaluated separately below. Proportions of different places from the three situations described in Section 5.2.1 regarding gazetteer entries are shown in Table 15.
Table 15: Proportions of places from the three situations from the input datasets.
Dataset Total nodes Anchor place Gazetteered Non-gazetteered
Campus 256 56 (21.9%) 143 (55.9%) 47 (18.4%) Melbourne 3520 974 (27.8%) 1684 (47.8%) 862 (24.5%)
In order to create ground-truth for both training the contextual- ized probabilistic search spaces as well as for evaluation, the loca- tions and granularities of all places from the two datasets are manu- ally grounded, and anchor places are linked to their corresponding gazetteer entries. 10-fold cross validation is used for training and eval- uating contextualized probabilistic-based search spaces. Specifically, triplets from the input PGDs are divided into 10-folds, and the search spaces for triplets in each fold are trained using ones from the rest based on the annotated granularity and location information. In the testing stage, the georeferencing procedure does not require further manual intervention. WordNet is used for creating dictionaries for determining the granularities of place from the test dataset.
5.3.2 Evaluation of DensityK
In this section, the proposed DensityK algorithm is compared to ex- isting clustering algorithms that are competitive for the task of place name disambiguation. There is a number of algorithmic features that will be focused on for discussion. The first one is robustness: that an algorithm should ideally work on different input datasets and have mimimum variance in precision and distance error. The next feature is minimum parameter-dependency. A parameter-free algo- rithm, or an algorithm with parameters automatically determinable, is desirable. Again, this is because for place name disambiguation, no pre-knowledge such as distances between places, or the extent of the space should be assumed for an input. Lastly, an algorithm should also ideally be parameter-insensitive, i.e., modifying parameter val- ues will not lead to significantly different results. Regarding these features, the degree of satisfaction of each of these algorithms will be discussed. An illustrative example input to the DensityK algorithm as well as other clustering algorithms to be compared is provided below in Figure 43 from the campus dataset. The ground-truth locations of these place names (the locations of their corresponding gazetteer en- tries), which are inside or near the University of Melbourne campus, are highlighted by red color in the bottom-right corner. For the algo-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.3 implementation and experiments 101 rithms to be tested below, each place name is considered as a suc- cessful disambiguation if it is correctly linked to its corresponding gazetteer entry.
Figure 43: An example input point cloud of all ambiguous gazetteer entry locations of a set of place names from the campus dataset, with ground-truth locations highlighted in red color.
5.3.2.1 Algorithms Evaluated for Comparison A total of sixteen algorithms are compared, including not only the ones that have been used for place name disambiguation be- fore, but also more generic clustering algorithms that exist in fields such as statistics, pattern recognition, and machine learning: over- all minimum distance (OMD), centroid, minimum distance to un- ambiguous referents (DTUR), DBSCAN, DBSCAN with automati- cally determined parameter (k-dist), OPTICS, OUTCLUST, CURE, CHAMELEON, HDBSCAN, KMeans, GMM, SNN, Spectral, SOM and DensityK. The first four algorithms have been used for place name disam- biguation before. The Overall minimum distance heuristic aims at se- lecting gazetteer entries so that they are as geographically close to each other as possible (Amitay et al., 2004; Habib and Keulen, 2012; Leidner, Sinclair, and Webber, 2003). The centroid based heuristic com- putes the geographic focus (centroid) of all ambiguous entry loca- tions and exclude entry locations that are too far away from it, pos- sibly in an iterative manner (Buscaldi and Rosso, 2008a; Smith and Crane, 2001). The Minimum distance to unambiguous referents heuristic identifies unambiguous place names, i.e., place names with only one gazetteer entry, and then disambiguate others based on minimum dis- tance to those unambiguous entry locations (Buscaldi and Magnini, 2010; Smith and Crane, 2001). The next four are density-based algorithms that leverage point den- sity for clustering. DBSCAN (Density Based Spatial Clustering of Ap- plications with Noise) relies on two parameters: the neighborhood distance threshold ε, and the minimum number of points to form a cluster MinPts. There is no straightforward way to fit the parame- ters without pre-knowledge of the data. A heuristic is proposed to
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 102 place georeferencing
estimate the value of parameters based on k-dist graph (a line plot representing the distances to the k-st nearest neighbor of each point) DBSCAN (Ester, M., Kriegel, H. P., Sander, J., & Xu, 1996). However, it is not trivial to detect the threshold, which requires a selection of value k as well as knowledge of the percentage of noise within the data. The algorithm OPTICS (Ordering Points To Identify the Cluster- ing Structure) (Ankerst et al., 1999) address the problem of parameter- dependency by building an augmented ordering of data which is con- 0 sistent with DBSCAN, but covers a spectrum of all different ε 6 ε. The OUTCLUST algorithm exploits local density to find clusters that are mostly deviating from the overall population (clustering by ex- ceptions) (Angiulli, 2006) given k, the number of nearest neighbors for computing local densities, as well as f, a frequency threshold, for detecting outliers. Hierarchical clustering algorithms typically build cluster hierar- chies and flexibly partition data at different granularity levels. The main disadvantage is the vagueness of when to terminate the iterative process of merging or dividing subclusters. CURE (Clustering Using REpresentatives) (Guha, Rastogi, and Shim, 1998) samples an input dataset and uses an agglomeration process to produce the requested number of clusters. CHAMELEON (Karypis, Han, and Kumar, 1999) leverages dynamic modelling method for cluster aggregation consid- ering k-nearest neighbor connectivity graph. HDBSCAN (Campello, Moulavi, and Sander, 2013) extends DBSCAN based on excluding border-points from the clusters and follows the definition of density- levels. Partitioning relocation clustering divides data into several subsets, and certain greedy heuristics are then used for iterative optimization. The KMeans algorithm (Hartigan and Wong, 1979) divides the data into k clusters through some random initial samples as well as an iterative process to update the centroids of the clusters until con- vergence. A Gaussian Mixture Model (GMM) (Celeux and Govaert, 1992) attempts to find a mixture of probability distributions that best model the input dataset through methods such as the Expectation- Maximization (EM) algorithm. KMeans is often regarded as a special case of GMM. There are other algorithms that do not belong to the previous three categories. The SNN (Shared Nearest Neighbours) algorithm (Ertöz, Steinbach, and Kumar, 2003) blends a density based approach by first constructing a linkage matrix representing the similarity, e.g., dis- tance, among shared nearest neighbors based on k-nearest neighbors (KNN). The remaining part of the algorithm is similar to DBSCAN. Spectral clustering relies on the eigenvalues of the similarity matrix (e.g., KNN) of the data and performs partition of the data into the required number of clusters. Compared to KMeans, spectral cluster- ing cares about connectivity instead of compactness (e.g., geometrical proximity). Kohen’s Self Organizing Maps (SOM) (Kohonen, 1998) is an artificial neural network-based clustering technique applying competitive learning using a grid of neurons. It is able to perform di-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.3 implementation and experiments 103 mensionality reduction and map high-dimensional data to (typically) two-dimensional representation. Different parameters of the algorithms are tested in a grid-search manner, as shown in Table 16. For k-dist, the author did not give a straightforward way to determine a threshold. Therefore, the 2σ rule in the same way as it is used in DensityK (Algorithm 3) is used, to enable a fair comparison. For algorithms that have not been used for place name disambiguation before (i.e., from k-dist to SOM), Algo- rithm 4 is used on the generated clusters for disambiguation. In case a top-cluster of a place name contains more than one gazetteer entries of this place name, the place name cannot be disambiguated and the case will be regarded as a failure.
Table 16: Parameter configurations of algorithms to be tested for place name disambiguation.
Parameter Notion Value Algorithms
Distance threshold (meters) ε 200, 2000, 20000 DBSCAN No. of nearest neighbors k 5, 10, 25 OUTCLUST, SNN, Chameleon, Spectral No. of clusters to derive c 3, 5, 10, 20 OPTICS, CURE, KMeans, GMM, Spectral Minimum points in cluster MinPts 1, 5, 10 DBSCAN, k-dist Frequency threshold f 0.1, 0.2, 0.5 OUTCLUST Weighting coefficient α 0.1, 1, 10 Chameleon SOM dimension m, n (5, 5), (10, 10), (20, 20) SOM
5.3.2.2 Results Table 17 presents the precision of each algorithm on the tested datasets, and the precisions are based on the best-performing pa- rameter configurations of these algorithms. Note that such best- performing parameter values may be difficult to determine without a-priori knowledge in real-world scenarios. DensityK achieves the highest precisions, followed by DBSCAN. This is not surprising, as DensityK is designed to be more flexible in determining cluster dis- tances compared to DBSCAN. In the remaining part of this section, the clustering results by each algorithm are discussed individually, in order to provide better insights of whether each of these algorithms is suitable for the task of this chapter, regarding both the feature re- quirements and performance. The clustering results generated by algorithms used for place name disambiguation in the literature, i.e., overall minimum distance, cen- troid, minimum distance to unambiguous referents, and DBSCAN, are shown in Figure 44, ranked by number of points contained. A major drawback of the overall minimum distance as well as the mini- mum distance to unambiguous referents methods is that they are sen- sitive to noise place names: place names with their actual location not captured by gazetteer. For example, the place name Union House is re- ferring to a building in the University of Melbourne campus. Its true location has no corresponding gazetteer entry, and the ambiguous gazetteer entries retrieved for this place name in the input point cloud are elsewhere around the world with the same name. Such cases are common for fine-grained place names, while prominent place names
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 104 place georeferencing
Table 17: Average precision of each algorithm with the best-performing pa- rameters on the tested datasets.
Category Algorithm Precision
OMD 76.7% Ad-hoc Centroid 57.2% DTUR 69.3%
DBSCAN 81.5% DBSCAN k-dist 75.4% Density-based OPTICS 73.2% OUTCLUST 70.6%
CURE 78.9% Hierarchical-based CHAMELEON 58.3% HDBSCAN 75.7%
KMeans 73.4% Partitioning relocation-based GMM 80.8%
SNN 70.5% Others Spectral 74.4% SOM 73.1%
The proposed algorithm DensityK 83.1%
(e.g., natural or political) are less likely to be missing in a gazetteer. Another disadvantage of the overall minimum distance method is scalability, as its time cost is significantly larger (over ten times) than other algorithms for most of the dataset tested, particularly for input with large number of place names and high ambiguities. The centroid- based method performs badly as the input point cloud is spread over the earth, and the centroid is somewhere in the middle and far from the actual focus of the ground-truth locations. DBSCAN is robust against noise place names, as it can capture the spatial context (the highlighted red region shown in Figure 43) of the original description and neglect entries outside of it. For the exam- ple point cloud, when the parameter ε is set to 2000m, the resulting disambiguation precision is higher than with other values selected from Table 16. More ground-truth entries are missed by the cluster generated with a value of 200m, and more ambiguous entries are included with a value of 20000m. For the clusters generated by the k- dist method, the value of ε determined in this case is roughly 300km, which is significantly larger than the most suitable value (somewhere between 1000 and 2000m). Consequently, k-dist performs badly in this case. Figure 45 shows clustering results generated by two other density- based clustering algorithms OPTICS and OUTCLUST for the exam- ple input data. OPTICS is designed to overcome the limitation of parameter-dependency of DBSCAN, thus it is expected to perform similar to DBSCAN with the best-performing parameters. The result shows that although OPTICS is more flexible in deriving clusters of
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.3 implementation and experiments 105
Figure 44: Clustering results generated by established clustering algorithms for place name disambiguation. various densities based on the tested datasets, this is actually a disad- vantage for the task of this chapter. OPTICS tends to aggregate points from the ground-truth spatial context with other points that are rel- atively close to it, despite that these marginal points have relatively larger local densities. In addition, the parameter NumberOfClusters (c) of OPTICS is problematic to define. Nevertheless, it is found that set- ting the value to 10 generally leads to optimal results regardless of input. OUTCLUST has the same drawback of merging nearby points from the spatial context, and it is decided by both parameters k and f. The two parameters are more sensitive to input data compared to c of OPTICS, and there is no straightforward method to determine the values either. A large input k value will result in few clusters, as more data points will be regarded as neighbors, and vice versa. Compared to OPTICS, OUTCLUST focuses more on relative density by consid- ering nearest neighbors rather than absolute density, thus, boundary points that are relatively close to some clusters while isolated from others, are more likely to be merged. Clustering results by hierarchical clustering algorithms are shown in Figure 46. CURE requires parameter c, similar to OPTICS. The
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 106 place georeferencing
Figure 45: Clustering results generated by density-based clustering algo- rithms.
Figure 46: Clustering results generated by hierarchical clustering algo- rithms.
derived clusters by CURE are generally similar to OPTICS. The al- gorithm CHAMELEON is more parameter-sensitive than CURE, and the resulting disambiguation precision is not as good as CURE even with best-performing parameter values. As for HDBSCAN, although it does not require any mandatory input parameters, the resulting precision for some input data is only slightly worse than DensityK.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.3 implementation and experiments 107
However, HDBSCAN is not robust against different input data – it performs quite well for some data, but significantly worse for others. This will be discussed in more detail later in this section. Clustering results for using the partitioning relocation-based algo- rithm are shown in Figure 47. The KMeans algorithm aims at min- imizing inter-cluster distances and dividing the data into k clusters. As a partition-based algorithm, it is not expected to perform well on fine-grained place name disambiguation, which is not a classifi- cation problem, and the resulting average precision is worse than HDBSCAN and CURE. For some input data, GMM performs well and achieves the same precisions as DensityK, or as DBSCAN with the best performing parameter values. The performance is generally good (measured by average precision) and robust (e.g., compared to HDBSCAN, which is discussed later). In addition, for most input data, setting different values of c, once larger than 10, makes little difference to the clustering compared to algorithms such as KMeans or CURE. Still, there is no easy way to automatically determine the value of c, and a single value does not always lead to the highest precisions for different input data.
Figure 47: Clustering results generated by partitioning relocation clustering algorithms.
Figure 48 shows the results using the remaining three algorithms. SNN is highly sensitive to the parameter k, the number of nearest neighbors to consider, and different k values often result in signifi- cantly different clustering results, as shown in the figure. A large k value tends to result in only a few large, well-separated clusters, and small local variations in density have little impact. Similar to OUT- CLUST, there is no easy way to determine a suitable, meaningful number of nearest neighbors to consider. Spectral clustering also has the problem of parameter sensitivity, both for c and k. Its precision is
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 108 place georeferencing
almost always worse than algorithms such as DBSCAN, CURE, and GMM, even with the best-performing parameter values. The result- ing clusters generated by SOM are often similar in pattern to those derived by CURE or KMeans, but the average precision is much lower (even lower than Spectral clustering). One advantage of SOM is that the SOM dimension can easily be set to large numbers, which typi- cally leads to higher precisions compared to adopting small values such as (5, 5). When it is set to more than (20, 20), continually increas- ing the values makes minimal difference to the resulting clusters, as well as precisions.
Figure 48: Clustering results generated by other clustering algorithms.
Finally, the result by DensityK is shown in Figure 49. The clusters generated are similar to DBSCAN with ε set to 2000m for this partic- ular input, as shown in Figure 44. Compared to the results generated by the other algorithms, as shown in Figure 45, 46, 47 and 48, it can be seen that the first-ranking cluster (the purple circles) generated by DensityK is most focused and similar to the highlighted ground-truth spatial context shown in Figure 43. From the tested algorithms, OPTICS, CURE, HDBSCAN, GMM, and DensityK seem to be most suitable for place name disambigua- tion considering the feature requirements. They provide good dis-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.3 implementation and experiments 109
Figure 49: Clustering results generated by the DensityK algorithm. ambiguation precision, and either do not require input parameters (HDBSCAN and DensityK), or have parameters easy to determine and work well on various input data (c for OPTICS, CURE, and GMM). In comparison, parameters such as k or ε are more sensitive to input, and cannot be determined easily each time a new input is given. Here the robustness of the five algorithms over different input data is further evaluated, in terms of variation in precision and aver- age distance error, i.e., the average distance between the ground-truth locations of place names and the entries selected by these algorithms. Documents are randomly selected from the dataset, and the results are shown in Figure 50. DensityK has constantly the highest preci- sion, as well as low variation compared to other algorithms such as OPTICS. In terms of distance errors, DensityK has the least variance of performance as well as overall minimum distance errors.
Figure 50: Disambiguation precisions (left), and average distance errors in km (right) by individual documents.
5.3.3 Results and Discussions of the Three-Step Georeferencing Approach
Figure 51 illustrates the procedure from initial input point clouds to the disambiguated anchor place locations after clustering. Precisions of disambiguation, as the result from the first step of the georeferenc- ing approach, are given in Table 18. Disambiguation failures in this step are due to three reasons. First, some place references are classified as anchor places but are actually
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 110 place georeferencing
Figure 51: Deriving cluster distance for input point clouds from the cam- pus graph (top) and the Melbourne graph (bottom); generating clusters for disambiguation from point clouds (left), and disam- biguated anchor places forming spatial contexts (right).
Table 18: Precisions of anchor place disambiguation.
Place graph Campus Melbourne
Precision 96.4%(54 out of 56) 91.9%(895 out of 974)
not. For example, Gate 4 is referring to an non-gazetteered place in the University of Melbourne campus but was identified as an anchor place, since there is a gazetteer entry with the same name and it was captured by clustering. Second, references to different places may be merged incorrectly by the graph merging approach (Kim, Vasardani, and Winter, 2017b), causing incorrect georeferencing of some refer- ences. For example, two buildings with similar references are both described to be near the same landmarks in the campus datasets, which are distinct places but are merged to the same node. In the dataset this only affects gazetteer matching (step three) of few places. Third, some anchor places are still ambiguous after the additional dis- ambiguation process, as no sufficient spatial relations are available for further disambiguation. For ALR derivation (the second step), the robustness of the train- ing approach is first tested by reducing the amount of training data and compare the results, in order to evaluate the sensitiveness of the approach regarding the amount of training data. A comparative ex- ample is presented in Figure 52. The figure on the left side is same as Figure 39 (a), while the figure on the right side shows the search space derived based on removing 80% of random training points. The
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.3 implementation and experiments 111 result indicates that even with a largely reduced amount of training data, meaningful and generally similar search spaces can be derived. Note that similarity is defined here by comparing to search spaces generated for other CCSs in terms of distributions by distance (c.f., Figure 53). Therefore, it is safe to conclude that the training approach is reasonably robust and does not rely on large amounts of training data in order to work.
Figure 52: Comparison of KDE-based search space generated by removing 80% of random input training points (right) with the search space generated with full training points (left) for testing the robustness of the training approach (distance [m])
Figure 53 gives examples of several trained search spaces. As shown in the three figures at the top, the area of the search spaces grows as the granularity of the relata becomes coarser, with other contextual factor values preserved. The search space for any relative direction is derived using training triplets with any of the four rela- tive direction relationships (in front of, left, right, back); since, for this particular case, this step is more interested to explore the metric dis- tance details through the generated search space instead of direction, as there is currently no reliable technique to automatically infer the reference frames and directions used in place descriptions. Some interesting observations can be made during the training stage. In the sample corpora, people tend to use certain relations under specific contexts. For example, prepositions expressing rela- tive directions are frequently used between building-level places, but rarely for places with granularities equal to or above district level. Prepositions expressing cardinal directions, on the other hand, are used more flexibly. Prepositions expressing qualitative distance re- lationships, such as near and at, are generally less frequently used when referring to places with granularities larger than street level. Topological relationships are typically used to describe relationships between places of different granularity levels (mostly inside). It is also noticed that the granularity of the relatum is the most influential con- textual factor on the shapes of search spaces in most of the contexts, followed by granularity of the locatum, spatial discourse granularity, and prominence. Some factors are more influential in certain contexts, e.g., the search spaces for near with prominent building-level places as relata are significantly larger than with less prominent ones.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 112 place georeferencing
Figure 53: Search space examples for triplets with building-level locata that have certain spatial relationships to relata from different levels, with prominence and spatial discourse granularity undetermined, generated by the KDE model (distance [m]).
The results of georeferencing using both formal and contextualized probabilistic-based search spaces are shown in Table 19. Three evalu- ation metrics are employed:
1. Precision: the percentage of places correctly linked to their cor- responding gazetteer entries.
2. ALR precision: the percentage of places with their corresponding gazetteer entries located within their derived ALRs.
3. Mean and median distance error: the mean and median dis- tances for all distances between the gazetteer entries matched and the ground-truth ones.
As shown in Table 19, the regression model performs best for all of the four evaluation metrics for both PGDs. The results given by the KDE and the tessellation model still perform better than the formal
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.3 implementation and experiments 113
Table 19: Georeferencing performance by search space models for the two input graphs
Graph Evaluation metrics Formal KDE Tessellation Regression
Campus Matching precision 36.3% 39.5% 39% 40.5% ALR precision 84.2% 93.2% 85.3% 94.7% Mean distance error 195m 153m 171m 144m Median distance error 96m 72m 88m 69m
Melbourne Matching precision 29.4% 32.5% 30.1% 34% ALR precision 73.8% 92.6% 85.5% 97.5% Mean distance error 7203m 5863m 6752m 5217m Median distance error 2410m 1675m 2250m 1469m model. Trade-off exists between (matching) precision and ALR preci- sion. This is because larger ALRs tend to result in higher ALR precision, but at the same time they are less restrictive, and thus matching preci- sion may be reduced. Therefore, mean and median distance errors are also used to provide additional information about to what spatial res- olution the derived ALRs are limiting the location of the places to be georeferenced. In summary, the ALR precisions show that most ALRs derived capture the location of the places to be georeferenced, while the distance errors are constraining enough considering the spatial resolution of the spatial contexts shown in Figure 51. When comparing the formal models and contextualized probabilis- tic models, the increases in ALR precisions are not simply because the areas of the new search spaces are larger. In fact, search space ar- eas for places that are finer in spatial granularity (e.g., building- and street-level) have generally decreased; yet most of them can still cap- ture the ground-truth locations of places to be georeferenced. Search spaces for places from coarser granularities, on the other hand, have generally increased and become able to capture more ground-truth locations. Thus, contextualized probabilistic based search spaces are more flexible to accommodate different contexts compared to the for- mal ones. Additionally, they provide likelihood distribution informa- tion which is useful for location visualization, particularly for non- gazetteered places. Figure 54 (a) shows matching precision according to overall score, i.e., the precision of places matched with score equal or greater than a threshold. Figure 54 (b) and (c) shows the distance errors for indi- vidual places from the two input graphs using different models. It can be seen that place references matched with similarities over 0.9 are generally around 90%. Overall, places matched with higher overall similarities are more likely to be correctly geo-referenced. Fig- ure 54 (b) and (c) plots the distance error for each individual place, to assist the interpretation of the previous provided mean and me- dian distance errors. There are some (relatively) small proportions of places with significantly larger distance errors than the other places, due to either incorrectly-georeferenced anchor places (error propaga- tion) or the lack of spatial relationship knowledge. For example, if a
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 114 place georeferencing
Figure 54: Precision by best-matching scores (a); distance errors by place for the two graphs (b, c) for both formal and regression-based model (distance [m]); precision and recall trade-off for identifying non- gazetteered places by thresholding (d).
place has no spatial relationship available to any anchor places, its ALR will be determined loosely by the whole spatial context, which could be significantly larger compared to other ALRs constrained by spatial relationships. In addition, an input PGD may include spatial relationships that are not true, either due to the imperfection of the parser, or mistakes by the descriptors. Still, such places can be located with a reasonable distance error, when comparing several kilometers as shown in Figure 54 to the whole area of the Melbourne dataset (Melbourne has a diameter of 120 km). In order to further understand how each of the three similarity measurement influences the best-matching result, a comparison of matching precisions when applying different weights of Equation 7 is provide in Table 20. The experiment is based on a grid-search of weights with 0.1 as change interval (except for the equal-weighted function as shown in the fist row). The previous results shown in Table 19 are based the best-performing weights. For the best-matching process, there are two reasons for failures. One is that some derived ALRs are not capturing the true locations of the corresponding places. The other is because some place refer- ences are too vernacular and different from their gazetteered names thus are challenging to be linked matched. Different weights of Equa- tion 7 for overall similarity measurement have been tested, as shown in Table 20. The result shows that string similarity generally plays the most important role in the matching process, while spatial simi- larity is least important. A likely reason is that the obtained gazetteer
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.3 implementation and experiments 115
Table 20: Matching precisions with different weights of the overall measure- ment function when applying the regression model
Overall measurement function W1 W2 W3 Matching precision
Equal weights for three similarities 0.33 0.33 0.33 25.3% Only string similarity 1.0 0.0 0.0 28.1% Only semantic similarity 0.0 1.0 0.0 12.3% Only spatial similarity 0.0 0.0 1.0 2.2% String and semantic similarity 0.5 0.5 0.0 22.5% String and spatial similarity 0.5 0.0 0.5 10.9% Semantic and spatial similarity 0.0 0.5 0.5 6.4% ...... Best performing weights 0.5 0.3 0.2 34.4% entries for each place to be matched have already been filtered by spa- tial relationship search spaces, thus string and semantic similarity are more effective for further ranking these entries than spatial similarity. Finally, non-gazetteered places are classified by testing different thresholds of best-matched scores, and the resulting precisions and recalls are shown in Figure 54 (d). Figure 55 provides an example of visualizing the approximate location of a non-gazetteered place swim- ming pool on the map by its ALR, given two relationships
Figure 55: Example of representing the location of swimming pool on map, given two spatial relationships to two anchor places. Search spaces of the two relations as contours (a); crisp ALR with 0.95 as threshold (b); crisp ALR with 0.5 as threshold (c); ground-truth location of the place (d) (Source: Google Maps, Jan 2018).
5.3.4 Comparison with Georeferencing Using a Basic Place Graph
A comparative experiment of the performance of the proposed geo- referencing approach tested on both the extended PGD model as well as the basic PG model introduced in Chapter 3 is conducted. There are two major differences between the implementations, as two types
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 116 place georeferencing
of knowledge are not captured in the basic model: places from the same discourse, and reference direction. The first difference is that spatial contexts are not separated in the basic PG for individual descriptions. An illustration is shown in Fig- ure 56 with two descriptions, each containing four anchor places. The spatial contexts of the two descriptions are different, as shown in the right-side figure (e.g., description one describes a street block in a city, and description two describes the whole city region). Dash lines in Figure 56 indicate spatial contexts, which are derived by the clus- tering algorithm based on the disambiguated locations of the anchor places. Using the basic PG model, there is no way to identify which places are originally from the same descriptions, and thus, only one spatial context can be derived, as shown in the left-side figure. Con- sequently, the ALRs generated for places in the first description will be inappropriately large if not further constrained by spatial relation- ships. In comparison, an extended PGD allows tracking places from the same descriptions, since the links are now preserved. Therefore, in this refinement, separated spatial contexts regarding each individ- ual description will be derived, as shown in the right-side figure. The refined spatial contexts will then be used for deriving ALRs, which are expected to be more constraining and thus more useful for georefer- encing.
Figure 56: The spatial context of a merged, original place graph (left), and separated spatial contexts of an extended place graph (right).
The second difference is that search spaces for relative direction relations can only be defined as buffered regions similar to near in- formation in a basic PG, as shown in Figure 57 (left). The reference direction of a relative direction relation can only be anchored, if the locations of the relatum and the place indicating reference direction (i.e., the place linked by the has_reference_direction edge) are available. The proposed refined search is illustrated in Figure 57 (right), with front indicating the reference direction, and the shaded regions repre- senting search spaces.
Figure 57: Search space of place B for relationship
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.3 implementation and experiments 117
Thus, in this experiment, places from input PGDs have been geo- referenced using four methods, namely baseline (using the basic PG model), SC (using the extended model, but only separates spatial con- texts for individual descriptions), RF (using the extended model, but only utilize reference direction knowledge), and hybrid (using the ex- tended model, separates spatial contexts for individual descriptions, and utilize reference direction knowledge). Refinement methods can only reduce the size of the ALR derived for a place. Thus, for each of the latter four refined georeferencing methods, four results are possi- ble when georeferencing a place:
• The size of the ALR is reduced compared to the one from the baseline, but both ALRs capture the ground-truth location of the place (Case 1).
• There is no change in the ALR’s size (Case 2).
• The ground-truth location is not captured in the either ALR (Case 3).
• The ground-truth location is captured by the ALR of the baseline method, but not in the reduced-size ALR (Case 4).
Figure 58 shows the percentages of places that belong to each of the four cases, grouped by the three georeferencing methods (the base- line is for comparison and therefore not shown). Places from Case 1 are regarded as better-georeferenced, while places from Case 2 and 3 are considered as equally georeferenced, and those from Case 4 are regarded as worse-georeferenced. Figure 58 shows that SC has the largest proportion of better georeferenced places, while for the RF method the percentages are much lower. In order to get refined ALRs for the RF method, relative direction relationships to some anchor places with reference direction information must be available. Since this is not always true, only part of the places to be georeferenced can benefit from these two refinement methods. When applying the RF method, some places are worse-georeferenced (Case 4). This is because some relative direction relation information is incorrect, ei- ther due to mistakes made by descriptors, or the imperfection of the reference direction annotation procedure. The hybrid method results in 90% of better-georeferenced places. This is expected as the method requires only information that can be used in any one of the three previous methods in order to make refine- ments. One drawback is that the hybrid method also has the largest worse-georeferenced place numbers due to error propagation. It is a trade-off problem between sizes of ALRs (the smaller the better as be- ing more constraining) and having ALRs capturing the ground-truth location of places. A measure of the refinement in terms reduced ALRs is depicted in Figure 59, which shows in percent the ALR remaining size after refinement compared to the baseline, for individual places. A value of 0.6, for instance, means the refined ALR is 60% of the size it was in the baseline method. Only places with the available required information for refinement are included in the figure. For example,
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 118 place georeferencing
Figure 58: Percentages of places from different ALR refinement situations compared to baseline.
Figure 59: ALR sizes maintained after refinement as percentages of the origi- nal (baseline) ALR size (y-axis) for individual places with refined ALR in decreasing refinement order (x-axis) using the SC, RF, and hybrid methods.
for the RF method, only places with relative direction relationships and reference direction information to some anchor places are in- cluded. The hybrid method results in the most size reduced ALRs for all places, which is also expected, given the method uses all the refining information available, combining the restrictions of the other methods. The distance errors between ground-truth and matched gazetteer locations for individual places are shown in Figure 60 for the baseline and the hybrid method, sorted by error size in the baseline. Large dis- tance errors in the baseline seem to be more likely to be reduced by the hybrid refinement method. A possible explanation is that large distance errors usually correspond to large, less restricting ALRs, and in such situations a refinement is more effective. On the other hand, if an ALR derived in the baseline is already constraining, further refine- ment might have no, or even negative effects. For example, the peak on the left side of the axis represents a place that was correctly linked
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 5.4 chapter discussion and summary 119 to its corresponding gazetteer entry in the baseline. In contrast, the refined ALR leaves out the ground-truth location, and causes the place to be miss-matched, resulting in an increased distance error.
Figure 60: Distance errors in meters between ground-truth and matched gazetteer locations for the baseline and hybrid methods for in- dividual places, ranked in increasing distance error order.
5.4 chapter discussion and summary
The objective of this chapter is to georeference all places in a PGD, regardless of whether they have been referred by gazetteered names or not. For this purpose, a three-step georeferencing approach is pro- posed. As the first step, a novel, robust, and parameter-independent al- gorithm DensityK is proposed for place name disambiguation. The algorithm is compared to competitive algorithms in terms of disam- biguation precision and distance error using the previous constructed tested datasets. These competitive algorithms are chosen from either ones that have been used for place name disambiguation in the liter- ature, or are from other communities (e.g., data mining) and are re- garded as promising for this task. Finally, the superiority of the new algorithm is confirmed as it, despite being parameter-independent, achieves state-of-art performance. This is reflected by that DensityK consistently achieves higher precision and has overall minimum dis- tance error compared to other competitive algorithms. The algorithm is also robust with different input data from various conversational contexts, as it has relatively low variance in disambiguation precision and distance error. It is able to derive clusters that are well-matched to the actual spatial contexts. As a density-based clustering algorithm, DensityK can be used for problems that other density-based algo- rithms (e.g., DBSCAN, OPTICS, and OUTCLUST) would be applied to. It is not suitable for classification problems, however, as part of the input data will likely be treated as noise. Furthermore, since DensityK aims at deriving clusters with significantly large point densities while not requiring manual-determined parameters, it is ideal for analysis
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 120 place georeferencing
of data with meaningful parts having distinct densities than the rest, particularly for data without a-priori knowledge. Georeferencing places without gazetteered references is the other main focus of this chapter, which has been largely ignored in the literature. Both formal search space models as well as several con- textualized probabilistic based ones trained from data are proposed, for various spatial relations from four different families: cardinal di- rection, relative direction, qualitative distance, and topology. The for- mal search space models are borrowed from fields including Artificial Intelligence and GIR. In comparison, the contextualized probabilistic- based search space models are novel as they are generalized, contex- tualized, and scalable. Multiple metrics have been used to evaluate these models, and the result shows that the contextualized probabilis- tic models are able to accommodate flexible contexts compared to the formal models, in terms of input with different sizes, spatial extents, and granularities. The method performs reasonably well in terms of precisions and distance errors, considering the spatial resolutions of the graph coverages as well as the novelty of the problem. Major ob- servations and failure cases, as well as their reasons, are discussed as well. In addition, a comparison of the performance of the proposed geo- referencing approach tested on both the extended PGD models as well as the basic PG model is given. Figure 58 shows that more than 90% of places are better georeferenced in terms of the quality of their de- rived ALR, and more than 60% of the ALRs are reduced in size with the hybrid refinement settings using the extended PGD model. The improvements are based on two types of information newly captured in the extended graph: places from the same discourse, and reference direction. Details of the reduced sizes of ALR as percentages and re- duced distance errors for individual places are presented as well. The results show that the improvements are particularly significant for certain places, such as ones that are from restrictive spatial contexts. The major limitation of this georeferencing approach is the rela- tively small training dataset for deriving the contextualized proba- bilistic search spaces, particularly under certain context criteria. Al- though the presented models are designed to be generalized enough and only require a small number of training samples, a richer train- ing dataset is still expected to further increase the georeferencing per- formance. Also, in this chapter only contextual factors that are au- tomatically obtainable from an input PGD are considered, while in the literature there are other factors identified that can affect the in- terpretation of spatial relationships as well, such as traveling mode and familiarity with the environment. In the future, the contextual- ized probabilistic search space models proposed in this work can be further refined by these factors, if the corresponding information can be obtained in some way. In addition, there is currently no link be- tween search spaces for different contexts, as they are trained using different data, even though the contexts may be similar. This results in difficulties in interpreting how the contextual factors affect search spaces of spatial relations.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] PLACEKNOWLEDGEQUERYING 6
A place graph database (PGD) provides abundant human knowledge about places extracted from place descriptions, including the relative locations of places described by place references and spatial relations, as well as other non-spatial and contextual information. Such knowl- edge is typically qualitative, and thus, a PGD can be regarded as a qualitative human knowledge base about places constructed from collective place descriptions. The captured knowledge is often selec- tive as the descriptors of the original input descriptions found worth describing, and the knowledge may be complementary to that pro- vided by the current geographic information systems (GISs) and spa- tial databases. A PGD should preferably support querying in order to utilize its captured knowledge in application scenarios. This chapter first analyzes types of queries in a categorical manner that can be answered by a PGD, while not all of these queries can be answered using the basic place graph (PG) model. Then, a framework for querying a PGD regarding these queries using multiple types of input is presented. Additionally, in order to contextualize query re- sults, a dialog-based approach is proposed for filtering query results through requesting additional user input. This chapter is based on content from a published paper (Chen et al., 2018) (introduced in Chapter 3). The chapter extends the discus- sions of queries answerable by the proposed PGD model. The query framework, which is only briefly introduced in the paper, is formal- ized in this chapter as well. This chapter additionally proposes a com- plementary method of querying a PGD using structured input, as well as an approach for query contextualization.
6.1 introduction
People frequently ask questions about places in their everyday lives (Harrison and Dourish, 1996), typically through either asking other people or submitting queries to web-based searching or question an- swering systems. Examples of questions about places include how to get to places, spatial relationships among places, as well as activities supported by places. The way people ask about place questions and the expected answers have been studied in several disciplines, includ- ing ones that focus on analyzing geographic-related web queries (Gan et al., 2008; Jones et al., 2008b) and ones from cognitive and psychol- ogy perspectives (Shanon, 1983; Winter and Freksa, 2012). With the increasing volume of unstructured text documents being published online (Melo and Martins, 2017), as well as the growing need for place-related information in everyday life, information systems that could provide richer and more locally valid data on place is desir-
121
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 122 place knowledge querying
able, as users might greatly benefit themselves by having such data in many situations (Davies et al., 2008). As a knowledge base with collective knowledge about places that people found worth describing, a PGD should support querying of the stored knowledge. Accordingly, the last task selected in this the- sis addresses place knowledge querying in a PGD. Querying a PGD is essentially based on graph traversals that match the stored knowl- edge according to criteria over the defined node types, properties, and edge connections of the PGD model. Therefore, this chapter first investigates place queries that are answerable by a PGD in a categor- ical manner. The diversity of properties of place as well as people’s need of place knowledge in their everyday lives have been studied in the literature, and this chapter will also link the proposed categoriza- tion to these previously identified properties and knowledge. The second subtask of this chapter is developing a querying frame- work. Considering the emergence of ubiquitous computing with spa- tial service users no longer necessarily be trained experts, it is prefer- able to have a query mechanism that requires minimal technical knowledge of users. A graph database is typically queried through graph traversal, yet it is challenging for non-expert users to formu- late such graph traversal queries. Therefore, this chapter proposes a complementary query approach addressing the user interaction with a PGD, enabling users to submit structured queries with the help of pre-defined dropdown lists. Structured queries will be translated into graph traversal algorithms automatically and conducted by a PGD. In addition, developing an approach for answering user queries tak- ing into account of context is necessary as well. Knowledge originally extracted from descriptions with different conversational contexts may have different degrees of relevance regarding a query, and thus the same query from different contexts may expect different results. A PGD allows certain levels of contexts of the extracted place knowl- edge to be captured, and therefore query results should ideally be filtered considering the context of queries if available. In comparison, queries by conventional graph traversal and exact-matching return re- sults aggregated from possibly multiple source contexts, which may be less useful compared to results already filtered by contexts. For this purpose this chapter proposes a dialog-based approach to present al- ternative contexts by interpreting the queries, and then to request additional refining information until the query can be proceeded. The remainder of this chapter is structured as follows. Section 6.2 addresses the three subtasks and provides a complete framework for querying a PGD. Implementation of the developed query framework is provided in Section 6.3 with two experiments. The first experiment demonstrates the query process as well as the superiority of the pro- posed PGD model over the basic PG model for the task of knowledge querying. The second experiment evaluates the contextualized query approach against the uncontextualized version. Section 6.4 summa- rizes and concludes the chapter.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 6.2 querying a place graph database 123
6.2 querying a place graph database
This section is divided into two parts. The first part lists and intro- duces types of queries that can be answered by a PGD. In the second part, the framework for querying a PGD is explained.
6.2.1 A Categorization of Place Knowledge Queries
The proposed PGD model in Chapter 3 has seven concepts modelled as types of nodes: place, reference to place, n-plet (a spatial relation- ship connecting places), formalized spatial relation, route (a sequence of places and spatial relationships), description (discourse), and user. Edges are defined for semantically connecting these concepts, such as a place must be referred by some place references, and place refer- ences and spatial relationships must be from n-plets. Queries of place knowledge in a PGD are based on searching for the stored nodes, edges, as well as properties associated with these nodes and edges that match some given criteria. As a graph-based database, the search operation is done by graph traversal. For instance, if a user wishes to find all café shops near Federation Square, a graph traversal opera- tion must be formulated to retrieve all places with type café and has near relationships to the place known as Federation Square. Based on the types of nodes, edges, and properties defined in the PGD model, a categorization of queries answerable by a PGD is pre- sented in Figure 61. The figure shows a hierarchical structure with each rectangle representing a type of query. There are three basic types of queries: spatial, non-spatial, and discourse. Spatial queries are about the location of places or the spatial relationships among places, while non-spatial queries are about the references, character- istics, and semantics of places. Discourse queries involve description- and user-level discourse information captured in a PGD that is not in- cluded by the other types of queries. For instance, a query that looks for all places described in any descriptions given by a user is cate- gorized as a discourse query. A query that has multiple components from different query types are regarded as a complex query in com- parison to a simple query. In the remaining part of this subsection, different queries are explained in detail.
6.2.1.1 Spatial Queries Spatial queries include queries about place location, spatial relation- ship, or route. Here place location is defined based on the georeferenc- ing information of places derived through the approach introduced in Chapter 5. Each place node is either stored with the location infor- mation of its corresponding gazetteer entry (typically as a point with coordinates), or an approximate location region (ALR) representing its approximate location. The georeferencing information is stored by the footprint property associated with place nodes. Accordingly, place location queries are interested in the footprint property values of place nodes to be matched.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 124 place knowledge querying
Figure 61: A categorization of types of queries that can be answered by a place graph database.
Compared to place location queries, spatial relationship queries are qualitative and can be further categorized by several situations, as shown in Table 21. In the table, lcm, rel, and rlm represents the lo- catum, spatial relation, and relatum components in an n-plet. For a query, a tick in the table indicates that the corresponding component is given by the query, while a question mark means the correspond- ing component is being queried. For instance, the query example in the first row “find Café near Federation Square” specifies the rela- tum Federation Square as well as the spatial relationship near and asks for the locatum. In a PGD, this query can be answered by a graph traversal operation that looks for places with a specific path connec- tion pattern. Other types of query in the table can be answered in similar ways. A PGD is able to answer all the query types from the ta- ble, although some queries are less frequently asked in people’s daily communications than others.
Table 21: Types of spatial queries about spatial relationships.
Query type lcm rel rlm Example query in natural lan- guage
Place query by spatial rela- ? XX "Find all Café near Federation tionship Square." Spatial relationship query X ? X "Where is Federation Square located in relation to Flinders Street Station?" Relatum query XX ? "Which suburb is Federation Square inside?" Location query X ? ? "Where is Federation Square located?" Environment query ? ? X "Tell me about other places re- lated to Federation Square." Spatial relationship instance ? X ? "What places are described as query near to each other?"
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 6.2 querying a place graph database 125
In addition to the previous two types of queries, a PGD also allows for querying routes that consist of places and their connecting spatial relationships occurring in sequential order. The stored routes in a PGD as route nodes can be queried in several ways, with examples including querying for places in a specific route, or querying for all routes that connect two specific places. Route knowledge is useful for applications such as navigation or identifying local landmarks.
6.2.1.2 Non-Spatial Queries In comparison to the previous spatial queries that focus on the spatial aspects of place knowledge, non-spatial queries are more interested in the referential, physical and functional aspects of places. Such non- spatial information of places is captured by the defined properties for place reference nodes in a PGD: place_reference, place_type, charac- teristic, equipment, affordance. Place references can be queried similarly as place location, regarding place reference nodes instead of place nodes. The physical aspects of places, such as physical features, elements, structure, and form, have been studied in the literature (Relph, 1976; Scannell and Gifford, 2010; Vasardani and Winter, 2016). In a PGD, these aspects are mainly reflected by the characteristic property. The value of the property can be a string describing the physical struc- ture and form of a place, e.g., ‘sandstone-made’, ‘symmetrical’, and it can also describes an emotional, subjective opinion of a place, e.g., ‘beautiful’, ‘peaceful’. Some of such values reflect the place proper- ties identified in the literature, such as local symmetries and good shape (Vasardani and Winter, 2016). The equipment property may also show the physical characteristic of a place. For instance, in the description “there is a statue in the hall”, the statue can be regarded as an object equipped by the place the hall, while it can also be regarded as a part of the physical structure of the place. The functional aspects of places are about the semantic categories of places (i.e., place type) as well as what activities and actions can be performed at these places (i.e., affordance). In a PGD, these aspects are reflected mainly by the type and affordance properties respectively. The equipment property value may also show the type or affordance of a place. In comparison to place type, place affordance reflects the semantics of a place with what human activities and behaviors can be associated to the place. Both place type and affordance are related to human experiences as meaning that people assign to a location, and thus, these properties may differ according to individual’s emotions, feelings, and social roles. In summary, a PGD capturing both the phys- ical as well as functional aspects of places can be used for querying the associated properties of places.
6.2.1.3 Discourse Query Up to now, there are two remaining types of nodes that have not yet been considered in the spatial and non-spatial queries introduced
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 126 place knowledge querying
above: description node and user node. When compared to the pre- vious queries which focus on the spatial and non-spatial properties of places, these two nodes provide discourse and therefore contex- tual information of these captured properties in a PGD extracted from place descriptions. Therefore, discourse query is selected as the last basic type of query other than the spatial and non-spatial ones in the categorization shown in Figure 61. Discourse queries are defined as queries that involves description or user nodes, either as criteria (e.g., find all places from a specific description) or query target (e.g., find descriptions that contain a spe- cific place). A simple discourse query matches description or user nodes by their properties, which does not involve places. Complex discourse queries, i.e., discourse queries involving nodes other than description and user nodes, on the other hand, are more likely to be conducted by PGD users that are interested about place knowledge. The spatial and non-spatial queries introduced above often expect results (e.g., place properties or spatial relationships) that may have to be related to the original discourses for analytical purposes. Com- plex discourse queries allows analytical queries to be performed by adding a discourse dimension to any spatial or non-spatial queries.
6.2.2 Query Framework
This subsection describes the framework for querying a PGD in two parts. Section 6.2.2.1 first introduces graph traversal as the typical way of querying a graph-based database. Then, querying a PGD using structured query input as a complementary approach is explained. Section 6.2.2.2 proposes a dialog-driven approach for query contextu- alization.
6.2.2.1 Querying Through Graph Traversal Graph databases are typically queried through graph traversal oper- ations, which can be done using defined query languages for some existing graph database platforms. In this thesis, the Cypher query language is used for explanation, as it is the embedded query lan- guage for Neo4j. In Neo4j, graph traversal is by default conducted in a depth first search (DFS) manner, although breadth first search (BFS) can be enforced as well. Cypher is functionally equivalent to SPARQL. Although their syntax is quite different, they share some similarities. For instance, they can both be used to select some data (using the MATCH clause for Cypher, and SELECT for SPARQL), and they can both be used for querying data in the form of triples. The RDF triple stores that SPARQL is used for and the labelled property graph model that Cypher is used for are both common tools in the semantic-web related studies for modelling triple data. A graph traversal operation matches nodes, edges, paths, as well as properties associated with nodes and paths according to the criteria given by a query. Such criteria functions as filters for graph traversal operations to search for particular patterns and properties. A query
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 6.2 querying a place graph database 127 such as “find computer labs that are inside the University of Mel- bourne” must be translated into a graph traversal operation in order to be conducted: finds all place nodes that are linked to place refer- ence nodes with their type property value as computer lab; the place nodes should also linked to some n-plet nodes, while these n-plet nodes are also connect to the inside spatial relationship node as well as the University of Melbourne node as relatum. The corresponding Cypher query is shown below, which is written in a way for better illustration and can be further simplified.
MATCH (pA:place)-->(prA:place_reference)-[{as:’locatum’}]->(np:nplet)<-[{as:’ relatum’}]-(prB:place_reference)<--(pB:place), (np:nplet)-->(r:spatial_relation) WHERE"computer lab" in pA.place _type and r.relation ="inside" and prB.place_reference ="University of Melbourne" RETURN DISTINCT pA
Formulating Cypher queries for a PGD requires expert knowledge and thus is difficult for non-expert users, as these users will have to be familiar with not only the syntax, but also the schema of the database in terms of nodes, edges, and properties. Therefore, this sec- tion proposes a complementary approach for querying a PGD using structured input that is more similar to natural language (NL) queries in its form. A query interface is proposed for users to select from pre-defined dropdown lists in order to compose a query, as shown in Figure 62. The interface supports configurations of four variables: tar- get (mandatory), criteria, order_by, and limit. The value of target can be a type of node or a property value. Criteria is a list of conditions to be applied as filters, and each condition can either be a property value or a connection to other nodes (e.g., has a relation to another place, being from a description, or being part of a route). The allowed prop- erties and connections are also pre-defined and are dynamic based on the selected target type. The Add criterion button opens a new window for choosing from properties and connections as well as for specify- ing values. Different criteria are separated by semicolons in the text box below the button that shows all the defined criteria. Order_by al- lows choosing from the target or one of the input criteria for result ranking, either in descending or ascending order. Limit is an integer for setting the number of target result entries to be displayed.
Figure 62: A complementary structured query interface with pre-defined dropdown lists.
Then, an automated process is developed for formalizing a query composed using the interface into a structured key-value pairs in the
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 128 place knowledge querying
format of JSON, and then into a Cypher query. The formated JSON of the example query above is shown below. The query interface allows all types of queries from Figure 61 to be answered. More examples will be provided in Section 6.3.1 during the implementation stage.
{ "target":"place" "criteria": { "place_type": ["computer lab"], "has_relation": [{"relation":"inside", "relatum":"University of Melbourne"}] } }
6.2.2.2 A Dialog-Driven Approach for Query Contextualization
Human place knowledge stored in a PGD often has context-dependent meanings. For instance, the two spatial relationships
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 6.3 implementation and experiments 129 results by the identified contexts. For illustration, given a query “what is the place on the right side of the Baillieu Library”, assuming the initial graph traversal operation generates two results: Brownless Biomedical Library, {reference direction: (South Lawn, as front)} Arts West, {reference direction: (South Lawn, as back)} The two places Brownless Biomedical Library and Arts West are the can- didate query results, while the remaining parts are the identified ref- erence direction contexts associated with these two place when they were described to be right to the Baillieu Library in the original de- scriptions. Next, a dialog will be initiated for requesting further user input for filtering by context: “Please select a reference direction - fac- ing the South Lawn or coming from the South Lawn?” If the user chooses facing the South Lawn, Arts West will be returned as the final result, while Brownless Biomedical Library is excluded as being incor- rect under the context specified by the user. The uncontextualized ver- sion of the query approach will return both places as query results aggregated from the two contexts, as they have both been described as having right relationships to the Baillieu Library. The dialog will iterate for a query if multiple types of contexts are identified. The complete procedure for querying a PGD is shown in Algorithm 5.
Algorithm 5 The procedure of querying a place graph database. Input: Query . A structured query or a Cypher query Output: QueryResult 1: if Query.isCypherQuery() = False then 2: Query := formalizeQuery(Query) 3: end if . Ensure an input query is in Cypher 4: 5: CandidateResults, ContextTypes := initialQuery(Query) 6: for ContextType in ContextTypes do 7: Contexts := getContexts(ContextType) 8: SelectedContext = initiateDiolog(Contexts) 9: CandidateResults := filter(CandidateResults, SelectedContext) 10: end for . Filter results by contexts 11: 12: return CandidateResults
6.3 implementation and experiments
This section describes the implementation of the query framework with two experiments. The first experiment provides demonstrative queries on the campus PGD using the query approach described in Section 6.2.2.1. The second experiment evaluates the contextualized query approach against the uncontextualized version based on ran- domly generated queries.
6.3.1 Query Demonstration
In order to demonstrate the superiority of the PGD model proposed in this thesis against the basic PG model, three query examples that can-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 130 place knowledge querying
not be answered with the basic model are selected. Note that all the queries that are answerable by the basic PG model are still supported. The three queries in NL are shown below:
• Find the most frequently referred relata (landmark query).
• Find places that are most frequently linked to the place Alice Hoy by spatial relations (place co-occurrence query).
• Find the most frequent paths of length three, consisting of only directional relationships, i.e., place A-relation a->place B-relation b->place C (prominent route query)
The corresponding structured queries (Figures 63, 64, and 65) and the translated Cypher queries of the three examples are shown below.
MATCH (p:place)-->(:place_reference)-[r{as:’relatum’}]->(:nplet) RETURN p, count(r) ASc ORDER BYc DESC
Figure 63: Structured query input for the first example.
MATCH (p:place)-[*2]->(n:nplet)<--({place_reference:’Alice Hoy’}) RETURN p, count(n) ASc ORDER BYc DESC
Figure 64: Structured query input for the second example.
MATCH path=(p1:place)-[*2]->(n1:nplet)<-[*2]-(p2:place)-[*2]->(n2:nplet)<-[*2]-(p 3:place), (n1)-->(:description)<--(n2),(n1)-->(r1:spatial _relation),(n2)-->(r2:spatial _ relation) WHERE r1.family in [’cardinal_direction’, ’relative_direction’] and r2.family in [’cardinal_direction’, ’relative_direction’] RETURN p1, r1, p2, r2, p3, count(path) ASc ORDER BYc DESC Table 22 shows the top results generated from the each of the three queries. The second column (results of the first query) shows the most
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 6.3 implementation and experiments 131
Figure 65: Structured query input for the third example. frequently referred relata as landmarks in the campus PGD. In a basic PG, landmarks can only be identified by counting unique relation- ships, as neither the number of occurrences of neither place refer- ences, nor instances of the same spatial relation between two places is preserved. For instance, the relationship
Table 22: Results for the three queries from the third experiment, with the second query using Alice Hoy as an example.
Rank Most frequent relata Places most fre- Most Frequent paths quently related to of length 3 Alice Hoy
1 University Of Melbourne Monash Road Old Arts, right, Baillieu Library, left, South Lawn 2 Union Building Entrance Baillieu Library, left, South Lawn, front, John Medley 3 Grattan Street University of Mel- Royal Parade, left, bourne Baillieu Library, left, South Lawn 4 South Lawn Wilson Hall Medical Building, left, Baillieu Library, left, South Lawn 5 Swanston Street Peter Hall Building Baillieu Library, left, South Lawn, left, Wil- son Hall
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 132 place knowledge querying
Figure 66: Database visualization of sequences of places connected by place reference and n-plet nodes returned by a query. Green: n-plet; Yellow: place reference; Purple: place.
6.3.2 Contextualized Querying
The proposed dialog-based query contextualization approach relies on the captured contextual information in a PGD for filtering the stored knowledge by context. It is expected that the approach is able to return query results that are more appropriate than results re- turned by the uncontextualized approach (without the dialog-driven process) regarding query contexts. This section first demonstrates the contextualized query approach. Then, an experiment for evaluating the query results returned by both approaches is conducted. The query pattern
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 6.3 implementation and experiments 133
> Different reference direction detected ======ID: 1: facing Old Arts (with Old Arts in the front) ======ID: 2: facing Union Building (with Union Building in the front) ======ID: 3: coming from Gate 10 (with Gate 10 at the back) ======ID: 4: coming from Grattan Street (with Grattan Street at the back) ======ID: 5: facing Baillieu Library (with Baillieu Library in the front) ======ID: 6: coming from Arts West (with Arts West at the back) ======> Please select a reference directions by id: 1 > Query result: Zoology, Botany, Baillieu Library, Arts West
In order to evaluate the contextualized approach, 500 random queries are generated for testing. The queries are generated by first randomly selecting places as relata from the campus PGD that have been referred by relative direction relations. Then, one of the associ- ated relative direction relations for each place is randomly selected. A generated test query consists of such a place as relatum and the selected relative direction relation. These queries are then conducted by both the contextualized and the uncontextualized approaches for comparison. For the contextualized query approach, when the system requests choosing one reference direction from the identified list for each query, a random one will be selected. The results are evaluated by precision and recall based on the groud-truth locations of places. An illustrative example is given be- low, with the same query example from above. The uncontextualized query approach returns eight places as results for the query. The con- textualized query approach randomly selects a reference direction with Baillieu Library in the front (facing the Baillieu Library) and re- turns two results. Based on the groud truth location information, each place from the query result can be validated. For instance, Arts West is a false positive in this case, since under the selected reference direc- tion it would sit on the right side of the South Lawn. Accordingly, the precision and recall for both query approaches for this query example can be computed, as shown below at the end.
Uncontextualized query result: Arts West, Baillieu Library, Botany, Co-Op Bookshop, Engineering Building Blocks, Grattan Street, Sciences Building, Zoology ======Initial candidates returned by contextualized query: [’Zoology’, ’Old Arts’, ’front’] [’Botany’, ’Old Arts’, ’front’] [’Baillieu Library’, ’Old Arts’, ’front’] [’Arts West’, ’Old Arts’, ’front’] [’Baillieu Library’, ’Union Building’, ’front’] [’Co-Op Bookshop’, ’Gate 10’, ’back’] [’Baillieu Library’, ’Gate 10’, ’back’]
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 134 place knowledge querying
[’Baillieu Library’, ’Grattan Street’, ’back’] [’Grattan Street’, ’Baillieu Library’, ’front’] [’Engineering Building Blocks’, ’Baillieu Library’, ’front’] [’Sciences Building’, ’Arts West’, ’back’] ——————– Randomly selected reference direction: Baillieu Library, front ——————– Contextualized query result: Grattan Street, Engineering Building Blocks ======Evaluation: Uncontextualized query precision: 0.25, recall: 1.00 Contextualized query precision: 1.00, recall: 1.00
Figure 67 shows the comparison of precisions for individual queries for both query approaches. The contextualized approach constantly achieves higher precisions than the uncontextualized ap- proach. This is expected, as the uncontextualized approach returns results aggregated from different reference direction contexts, and some of these results may be wrong regarding the reference direc- tion selected. The fact that the contextualized query approach does not always achieve the precision of 1 is due to the incorrectness of some stored knowledge. Such incorrect knowledge is either caused by mistakes made by the descriptors, or due to the fact that the anno- tated reference direction information in the campus PGD is not always correct. Such incorrect knowledge is not manually excluded before the experiment, thus the precision results for both cases when the stored knowledge is correct and partially incorrect are shown and can be compared. It can be seen that even for the latter case (the stored knowledge is partially incorrect), the resulting precision for the contextualized approach is still equal or higher than the uncon- textualized approach.
Figure 67: Precision comparison for individual queries for the uncontextual- ized and contextualized query approach.
Some randomly generated queries have only one reference direc- tion context identified among the candidate results. In such situations
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 6.4 chapter discussion and summary 135 the results from both approaches will be the same. Figure 68 further shows precisions for queries that have multiple contexts identified. It is observed that for some queries both query approaches still result in the same precision. This is because sometimes even multiple reference direction contexts have been identified for a query, candidate results from contexts different than the selected one are still valid based on the ground-truth information, in terms of having the relationship to the relatum in the query under the selected reference direction.
Figure 68: Precision comparison considering only queries with multiple con- texts identified.
The recall comparison is shown in Figure 69. The recall of the un- contextualized approach is always 1 as it matches spatial relation- ships regardless of contexts, and the result of a query using the con- textualized approach is always a subset of the result using the uncon- textualized approach. The reasons that the recall of the contextualized approach is not always 1 are due to incorrect stored knowledge, as well as situations where candidate results from unselected reference direction contexts are still valid, as both explained above.
6.4 chapter discussion and summary
The chapter begins with an analysis of queries that can be answered by a PGD, which can be categorized by three basic types: queries about place locations, qualitative spatial relationships, and routes (spatial queries), queries about place references, characteristics, and seman- tics (non-spatial queries), and queries that involve description- and user-level knowledge (discourse query). The identified query types are also linked to previous research about human place knowledge and place properties, showing that a PGD is capable of capturing this knowledge as well as place properties while allowing them to be queried. Then, this chapter describes the framework of querying a PGD based on graph traversal operations. Particularly, an approach for querying a PGD using structured input is developed. It is a comple-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 136 place knowledge querying
Figure 69: Recall comparison for individual queries for the uncontextual- ized and contextualized query approach.
mentary query approach compared to conventional graph database traversal query languages (e.g., Cypher). The purpose is to minimize the technical effort for querying a PGD, especially for non-expert users. The approach also provides a jumping point for future research on querying a PGD using NL input, as NL queries may be formalized as structured queries using parsers. The approach allows for all types of queries identified previously to be queried in a PGD. Additionally, in order to contextualize query results, this chapter proposes a dialog- driven approach for requesting further user input of contextual infor- mation and filtering query results accordingly. The approach captures and uses of the user’s context and compare to the context of knowl- edge stored in a PGD in order to generate filtered results. The proposed query framework is implemented and tested with two experiments using the campus dataset introduced in Chapter 3. The first experiment demonstrates the proposed query framework using several query examples, showing that different queries can be formalized into graph traversal algorithms and conducted by a PGD. The selected queries also show the superiority of the PGD model de- veloped in this thesis, as these queries cannot be answered using the basic PG model proposed in the literature. The second experiment evaluates the contextualized query approach and compares it with the uncontextualized version over randomly generated queries. The result shows that the contextualized approach consistently achieves equal or higher precisions for the generated results, particularly for queries with results associated with multiple contexts. The result also shows that the contextualized approach achieves 1 recall for the vast majority of tested queries, expect for situations caused by incorrect or incomplete knowledge captured in the test PGD. One major limitation of the implemented contextualized approach is that it does not further utilize relational reasoning mechanisms for query expansion, and it currently only relies on exact matching. With a reasoning mechanism for query expansion, the resulting recall is
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 6.4 chapter discussion and summary 137 expected to further increase. For instance, reference directions cap- tured as (Gate 10, as_back) and (Old Quad, as_front) may refer to actually the same direction, and this fact may be identified through a reasoning mechanism. However, the current matching process will only match either one of the two reference directions according to the user input and ignore the other one. One future solution is to re- trieve all relative direction relationships from the database, which are connected to the relatum place from a query, instead of only the rela- tionship specified in the query. Then, a reference direction reasoning mechanism can be developed to exclude candidates that are contra- dicting with the query context. In this way, the returned results will be a superset of the results returned without the reasoning mecha- nism. Such a query expansion approach can be extended to support other types of contextual factors as well with additionally defined reasoning rules. While the previous three tasks of this thesis (i.e., modelling, rea- soning, and georeferencing) focus on knowledge modelling and pro- cessing, querying a PGD concerns knowledge output and utilization. A PGD is expected to support querying of its stored knowledge in order to be regarded as a human place knowledge base. This chapter has demonstrated the capacity of a PGD for answering different types of queries about place knowledge using the developed query frame- work, as well as its capacity to process context and return contextual- ized query results with additional user input. As a knowledge base, a PGD is often highly selective and incomplete compared to traditional spatial databases, collecting only facts that people found worth to ex- press in some conversational contexts. Traditional spatial databases are committed to maintain a complete record of geometric represen- tations of the space, while completeness is not the aim of a PGD. The PGD model is developed based on the open world assumption and aims at providing a data structure for human place knowledge repre- sentation, as well as providing complementary knowledge to current GISs and spatial databases with NL as a new type of input.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] [ April 29, 2019 at 20:02 – classicthesis version 5 ] PLACEGRAPHDATABASEMANAGEMENTSYSTEM 7
This chapter describes the design and implementation of the place graph database (PGD) management system. The system is modular and implements the major contributions of this thesis, i.e., approaches introduced in the previous chapters for the modelling, reasoning, georeferencing, and querying tasks. The system is additionally able to perform tasks including description parsing, graph creation from parsed files, database visualization, and place mapping, thus provid- ing a processing chain for manipulating human knowledge about places extracted from place descriptions. This chapter first provides an overview of the system, including the interface layout and the in- tegrated modules. Then, each module will be explained in detail in the following sections.
7.1 system overview
Figure 70 shows the interface of the developed system. The interface is divided into four zones: the module tabs zone, the textual output zone, the database visualization and manipulation zone, and the map- ping zone. The module tabs zone contains six tabs, corresponding to six major functional modules of the system: database management, input preparation, database construction, georeferencing, place map- ping, and querying. The textual output zone shows the textual results and notifications of any database operations. The database visualiza- tion and manipulation zone is built using the Neo4j application pro- gramming interface (API), and it is able to visualize a PGD by color- coded nodes and edges. The zone also allows raw Cypher queries to be executed in order to show/hide nodes, edges, and properties for different visualization views. Finally, the mapping zone is used by the input preparation and place mapping modules for selecting an area on the map, and map the locations of georeferenced places respectively. Figure 71 illustrates the dependency relationships among the six modules as well as their main functions. A dependency relationship (indicated by an arrow) means a module is dependent on the other module in order to execute some or all of its functions. For example, the place mapping module requires places to be georeferenced first in order to map them, thus it has a dependent relationship to the georeferencing module. The mapping module is also dependent on the querying module in order to filter places to be mapped by queries. In the remaining part of the chapter, the functions and individual interfaces of each module will be introduced.
139
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 140 placegraphdatabasemanagementsystem
Figure 70: Interface layout of the developed place graph database manage- ment system.
Figure 71: Dependence relationships among the six functional modules and their main functions.
7.2 database management module
The database management module allows database configurations, mainly for setting the directory of the PGD to be created or opened. The buttons to enable and disable the database visualization and ma- nipulation zone are also under this tab. Figure 72 shows the interface of the database management module in the module tab zone.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 7.3 input preparation module 141
Figure 72: Interface of the database management module in the module tab zone.
7.3 input preparation module
Currently there are four input sources for constructing or populating a PGD. The first and second sources have already been introduced in Section 3.3: through asking people to give descriptions about places (e.g., the University of Melbourne campus dataset), or harvesting de- scriptions as textual documents from websites using some already developed techniques (e.g., the Melbourne dataset) (Kim, Vasardani, and Winter, 2015). The raw place descriptions collected through these ways have to be parsed before being used for PGD construction. The third source is a web-based treasure-hunting game that can be accessed through https://placegame.eng.unimelb.edu.au/. People registered in the game will see random coins displayed on the map, as shown in Figure 73. Then, they will have to collect these coins by going to their corresponding locations on the map and give descrip- tions about the surrounding environments at these locations.
Figure 73: A web-based game for collecting place descriptions.
The last source is based on a developed method within this PhD project for deriving triplets from OpenStreetMap (OSM) (Hamzei et
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 142 placegraphdatabasemanagementsystem
al., 2018), with partial contributions including problem conceptualiza- tion, methodology, implementation, and experiment. The workflow is illustrated in Figure 74. The method has been integrated in the devel- oped system as well, although it is not taking place descriptions as input.
Figure 74: Workflow of generating place graphs from maps.
The input preparation module has two sub-modules, and it is used for pre-processing input from these four sources into formalized files that can be directly used for PGD construction under the database con- struction tab. The two sub-tabs are shown in Figure 75. The first tab integrates the parser for parsing place descriptions (Liu, Vasardani, and Baldwin, 2014), and it will store the parsed data as formalized files for database construction. The process of parsing descriptions has been introduced in Section 3.3 with an example given. The sec- ond tab implements the method for deriving triplets from OSM.
Figure 75: Interface of the input preparation module in the module tab zone and its sub-tabs.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 7.4 database construction module 143
7.4 database construction module
The database construction module tab is shown in Figure 76. The module is used for constructing a PGD automatically from a file pro- duced using the previous input preparation module. It allows two types of input: csv and JSON. A csv file (as a set of triplets) is used for constructing a database of a basic place graph (PG) (Vasardani et al., 2013), and a JSON file is for constructing an (extended) PGD. The reason to implement the function for constructing basic PGDs in the system is because some experiments in the previous chapters of this thesis require comparison of the results delivered by the two types of PGDs. The function may also be useful for future studies and therefore is kept.
Figure 76: Interface of the database construction module in the module tab zone.
An example of constructing a PGD from an input file has been given in Section 3.3. During the creation or population of a database (af- ter selecting the file to import and clicking the import data button in Figure 76), the consistency-checking approach, as introduced in Chapter 4, is conducted. Specifically, it will infer unknown spatial re- lationships over cycles and detect inconsistencies. Inconsistent spatial relationships will be flagged in the textual output zone for notifying the user. After construction of a database, a user is able to view the database in the database visualization and manipulation zone.
7.5 georeferencing module
The tab of the georeferencing module is shown in Figure 77. The three buttons at the top are corresponding to the three steps of the georef- erencing approach introduced in Chapter 5, i.e., disambiguation of anchor places, deriving approximate location regions (ALRs) for non- anchor places, and gazetteer best-matching of non-anchor places. Ei- ther formal search space or trained probabilistic-based search space can be selected for georeferencing. Another button at the bottom right hand side allows processing the three steps all together. Once each step is initiated, the system will georeference places that has not been georeferenced yet (e.g., for places in a newly created database, or in an existing database which has been populated with new data) and ignore places that has already been georeferenced.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 144 placegraphdatabasemanagementsystem
Figure 77: Interface of the georeferencing module in the module tab zone.
Example georeferencing textual output from each of the three steps shown in the textual output zone is given in Figure 78. While the textual output is being printed in the zone, the database is being updated and the georeferencing information is being added to the corresponding place nodes.
Figure 78: Georeferencing results from each of the three steps shown in the textual output zone: disambiguation of anchor places (top), deriv- ing approximate location regions for non-anchor places (middle), and gazetteer best-matching of non-anchor places (bottom).
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 7.6 place mapping module 145
7.6 place mapping module
The mapping module allows places in a PGD that have already been georeferenced to be mapped in the mapping zone. The interface of the module tab is shown in Figure 79 (top). The left-hand side is for configuration and selection of which places to show. Specifically, it allows to show either anchor places or non-anchor places, as well as to show either ALRs or best matched gazetteer entry of each place on the map (the latter one is only for non-anchor places since anchor places do not have ALRs). The right-hand side is for displaying a table of places associated with their georeferencing information, and a user can select places to be mapped from the table. Figure 79 (bottom) shows an example of two places (Electrical & Electronic Engineering and Sydney Myer Bldg) being mapped in the mapping zone with both their ALRs and matched locations shown. The ALRs are shown by crisp boundaries.
Figure 79: Interface of the place mapping module in the module tab zone (top), and an example of two places being mapped in the map- ping zone with both their approximate location regionss and matched gazetteer entry locations (bottom).
Other than relying on pre-defined criteria and table for mapping as illustrated above, the system additionally allows filtering by query under the querying tab. Details are given in the next section.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 146 placegraphdatabasemanagementsystem
7.7 querying module
The query module has two sub-tabs. The querying framework intro- duced in Chapter 6 has been implemented with a query interface shown in Figure 80 (top). The second tab allows for raw Cypher queries to be executed for expert users. Several query examples have already been given in Section 6.3.1.
Figure 80: Interface of the querying module in the module tab zone and its sub-tabs.
The first sub-tab allows a user to select from drop-down menus of pre-defined query components, which will then be translated into a Cypher query by the system automatically and executed. It also allows a user to submit the query, and select whether to contextualize the query or not. If a query is contextualized, then the user may have to further specify context criteria using the textual input field in the middle. Figure 81 provides an example of the textual output of a contextualized query.
Figure 81: An example of contextualized query through a dialog.
If a query is conducted and some places nodes are returned, the system allows mapping of the returned places similarly as illustrated in the last section. There are two bottoms for mapping the returned place in each of the two sub-tabs respectively, as shown in Figure 80.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] DISCUSSIONANDCONCLUSIONS 8
Place, as a human spatial concept used by people in cognitive and communicative tasks, is not immediately representable by the geographic information systems (GISs) and spatial databases de- veloped over the past half-century. While these GISs and spatial databases are typically based on unambiguous, crisp, and metric ge- ometries and coordinates as the general paradigm of space represen- tation, places are often spatially vague, infused with human experi- ence, and have context-dependent meanings. The fundamental mis- match hinders the development of information systems that capture and reflect the way people think and reason about space, which can be leveraged for smoothing and simplifying interactions between hu- man and computers. The need for such systems is growing, facing the emergence of ubiquitous computing with users no longer necessarily be trained GIS experts. However, despite the recognized importance of place-based research in geographic information science (GIScience), not much success has been made and no widely adopted computa- tional data model for place has been developed. Place descriptions reflect the way that people mentally repre- sent and communicate about place knowledge. The abundance of natural language (NL) place descriptions suggests complementary place-based approaches to human knowledge representation. The web provides a plethora of place descriptions in various forms, and thus the web offers a unique opportunity for the collection of rich place description datasets. Towards the goal of developing a place- based GIS, this thesis aims at contributing to several research aspects in order to identify and minimize gaps in knowledge. Regarding place descriptions as a new type of data source that is unstructured and not directly supported by current GISs and spatial databases, this thesis has addressed challenges of modelling, reasoning, georeferenc- ing, and querying with information extracted from place descriptions using a place graph database (PGD). The contribution of this thesis is twofold. The first fold includes developed approaches addressing these four tasks as scientific contri- butions to knowledge: a) the understanding of place-related knowl- edge embedded in NL place descriptions, and ways to model them, b) reasoning with the extracted qualitative spatial relationships as hu- man spatial knowledge of places, c) locating places from descriptions through disambiguation, spatial relationship interpretation, and con- textualization, and d) identification of types of place knowledge that can be queried in a PGD, as well as ways to query them. The second fold is an implemented system that integrates functions correspond- ing to the four tasks. The database provides selective and collective information about places given by people, and such information may be complementary to the information that traditional GISs and spatial
147
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 148 discussion and conclusions
databases provide. With such a system as a human knowledge base about places developed, it would be possible to further explore how this system could be used for assisting and improving current ser- vices that involves human-computer interaction. The ultimate goal of such a system is the spatially intelligent machine proposed by Winter and Wu (2009) that is capable of understanding and producing hu- man spatial language, thus allowing smooth communication between human and machine. The remaining part of this chapter is organized as follows. Sec- tion 8.1 presents the major results and findings for the four major challenges, according to the contents in Chapters 3, 4, 5, and 6 re- spectively. Section 8.2 summarizes the core contributions of this the- sis and evaluate them against the proposed hypotheses in Chapter 1. Section 8.3 identifies the limitations of this thesis, and discusses di- rections for future research.
8.1 discussions of the major results
In this section, summaries of the developed approaches regarding the four major research tasks, along with discussions of the obtained re- sults are given. The detailed contributions and significance of these developed approaches, as well as their future applicability are dis- cussed as well.
8.1.1 Place Knowledge Modelling
Place descriptions provide a rich source of human knowledge about places. Place references and their mutual qualitative spatial relation- ships from place descriptions have been used to construct place graphs (PGs) in the literature (Vasardani et al., 2013). However, sev- eral limitations of this model have been identified in Section 3.2. For instance, the triplets used to build such a basic PG are stripped off of much of their conversational contexts, and the model considers only limited types of knowledge embedded in place descriptions. Con- sequently, the first task of this thesis is to design a data model for capturing information extracted from place descriptions. The model should also allow efficient storage as well as retrieval of the captured knowledge for the later three tasks of this thesis. Chapter 3 largely reorganizes, revises and extends the basic PG model from the literature. It does not only capture information from place descriptions that is useful but ignored by the basic model, but also revises the model conceptually. Eight types of information that are embedded in place descriptions and not captured in the basic PG model are identified in the first place: place semantics and character- istics, places and spatial relations from the same discourse, as well as their sequential order of appearance, reference frame, non-binary relationships, co-occurrence of place references and spatial relations, place conceptualization, route and accessibility, and description con- text and source context. The model consists of seven types of nodes
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 8.1 discussions of the major results 149 representing conceptual entities: place reference, n-plet, place, route, spa- tial relation, description, and user (instead of only one place node type in the basic model) as well as nine types of edges representing the semantic connections between these conceptual entities. Each type of node or edge has associated properties representing their attributes. The model is implemented using a commercial graph database plat- form Neo4j with two description datasets. A database management system interface capable of various PGD operations has been devel- oped as well. The database supports structured JSON with parsed in- formation from place descriptions as input, and allows efficient query through graph traversal. The superiority of the model compared to the basic PG model is demonstrated in Chapters 4, 5, and 6. Specifi- cally, for the reasoning task, inference and consistency reasoning for relative direction relationships cannot be performed without refer- ence frame and reference direction information being captured; for the georeferencing task, the performances of the developed approach leveraging the extended model and the basic model are compared, and the advantage of the extended model is confirmed; for the query- ing task, various types of queries have been identified that cannot be conducted using the basic model. The extended PGD model captures spatial, semantic, and contextual information about places from place descriptions, and such informa- tion can be linked to the five core concepts of spatial information proposed by Kuhn (2012): location, field, object, network, and event, as already discussed in Section 3.4. The model allows extensions con- sidering future needs for modelling other elements from spatial lan- guage. For example, although locomotions such as walk and go, which are common in route descriptions (Winter et al., 2018), are currently not an element in the model, they can easily be modelled through introducing more conceptual entities and relationships as node and edges associated with properties. In addition, the knowledge mod- elled in the extended PGD model can help better understand human descriptions.
8.1.2 Spatial Relationship Reasoning
Knowledge contained by place descriptions are typically qualitative, in which people use various qualitative spatial relations to describe the relative locations of places within some environments. A PGD cap- tures this knowledge and thus, spatial relationships stored in a PGD are used for relational inference and consistency maintenance as the second task. The motivation is two-fold: a PGD is often incomplete in terms of spatial relationship knowledge, therefore a method that can infer spatial relationships from existing ones is helpful by pro- viding a more complete graph, as tasks including georeferencing and querying require as much relationship knowledge as possible. Also, in order to maintain and query a large PGD, mechanisms to maintain relational consistency of the knowledge stored during transactions is
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 150 discussion and conclusions
necessary. The correctness of the knowledge stored in a PGD is also crucial for tasks including georeferencing and querying. Chapter 4 proposes a flexible framework for relational inference as well as for maintaining relational consistency in a PGD. For this pur- pose the chapter suggests relational composition rules (for each sin- gle family, as well as for all spatial relationships together) as reason- ing constraints and a path-based algorithm for consistency checking. Four types of spatial relationship families are considered, building on the established field of qualitative spatial reasoning (QSR): cardinal direction, relative direction, qualitative distance, and topology. The framework first deals with spatial relations from each family, and it is then extended to support cross family reasoning. As an outcome, the framework is able to identify and flag inconsistent relationships within a limited path length. It has no further mechanisms to de- cide which relationship is truer or closer to the representation of geo- graphic reality. The experimental results show that the implemented system of the reasoning framework is robust and computationally feasible within its pragmatically limited query depth. For triplets both from the cam- pus description datasets as well as the manually made set, inconsis- tent relationships were flagged and consistent ones were accepted. False positives were actually due to a more flexible use of language than this system was committing to. Therefore, even with an imper- fect mapping between NL spatial relationship expressions and formal spatial relationships, as well as a limited query depth, the framework provided robust evidence of working correctly. For each triplet to be created or updated, the database tests consistency of the new relation- ship over any limited cycle that will be formed by the added triplet. Thus whenever a transaction is being processed, the place database will either end in a locally consistent state, or flag the new triplet for human inspection. The chapter provides a first step into using a graph database to reason with human qualitative spatial relationship knowledge about places extracted from NL place descriptions, as well as to build a rela- tionally consistent database for storing such knowledge. Such a PGD, built and maintained in relational consistency, can then be more re- liable for relevant applications (e.g., navigation) compared to a PGD without such a mechanism. The reasoning framework proposed al- lows convenient future extensions of new spatial relations and con- text factors through adding associated reasoning rules. Finally, this thesis provides insights of cross family reasoning for future research based on observations from preliminary experiments. Other relations, including non-binary ones such as between and through, are more chal- lenging to be applied for reasoning. Discussions also include the lim- itations of logical approaches due to loss of context and mismatches between NL pragmatics and the axioms of logical calculi. Future work must address the current weaknesses experienced in this paper, which requires an even deeper consideration of context. The usage as well as interpretations of spatial relationships in NL is flexible and context-dependent. While consistency rules can be set out from a log-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 8.1 discussions of the major results 151 ical perspective, such rules do not always reflect pragmatics from a linguistic perspective.
8.1.3 Place Georeferencing
Place descriptions with references to places and spatial relationships provide spatial knowledge of places in their relative locations, and such knowledge has already been captured in the proposed PGD model. The third task of this thesis is identified as georeferencing all places in a PGD, based on the captured references to places as well as their mutual spatial relationships. The georeferenced places can then be linked to other GISs, spatial databases, and spatial services for application scenarios. The problem is more challenging in com- parison to place name disambiguation, as one major task in the field of toponym resolution (TR). This is because place descriptions are of- ten flexible, vernacular, and contain names referring to fine-grained places, places not referred to by official names, as well as places not captured in official databases such as gazetteers. Places from these situations are challenging to resolve and require different methods than previous proposed conventional ones for TR. Chapter 5 introduces a three-step georeferencing approach for geo- referencing all places in a PGD, regardless of whether they are re- ferred to by gazetteered names or not. The first step attempts to georeference places that have been referred to by gazetteered names. For this purpose, a novel, and parameter-independent, and robust clustering-based disambiguation algorithm is developed. It does not require manual input parameters and is flexible for disambiguating places from place descriptions with different conversational contexts. The second step leverages spatial relation search space models on the stored spatial relationship between the remaining places and places resolved in the first step, in order to derive approximate location re- gions (ALRs) for these remaining places. The search spaces models include both formal ones as well as contextualized ones trained from data. Finally, a weighted multi-value similarity measuring approach is presented for matching places from the second step to gazetteer entries, based on their derived ALRs as spatial constrains. Even if the matching fails, i.e., the place is non-gazetteered, the derived ALR from the second step can still be used to visualize the location of this place on a map. The novel clustering algorithm is compared to competitive algo- rithms in terms of disambiguation precision and distance error us- ing the previous constructed tested datasets. These competitive al- gorithms are chosen from either ones that have been used for place name disambiguation in the literature, or are from other communi- ties (e.g., data mining) and are regarded as promising for the task of place name disambiguation. Finally, the superiority of the new algorithm is confirmed as it, despite being parameter-independent, achieves state-of-art performance. For the second and third step, mul- tiple metrics have been used to evaluate these models, and the result
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 152 discussion and conclusions
shows that the contextualized probabilistic models are able to accom- modate flexible contexts compared to the formal models for input with different sizes, spatial extents, and granularities. The method performs reasonably well in terms of precisions and distance errors, considering the spatial resolutions of the graph coverages as well as the novelty of the problem. Major observations and failure cases, as well as their reasons, are discussed as well. In addition, a comparison of the performance of the proposed georeferencing approach tested on both the extended PGD models as well as the basic PG model is given. The result shows that the extended model allows most places better georeferenced in terms of the quality of their derived ALRs, and the improvements are particularly significant for certain places. The georeferencing approach proposed enables mapping of places in a PGD, as well as makes it possible to establish links to coordi- nates in GISs and spatial database. The contributions of this phase is three-fold aiming at resolving places from different cases, consider- ing the mapping between place and place reference. Accordingly, the georeferencing result can either be an entry from a gazetteer with a geometrical representation (typically a point) for places from the first and second cases, or a density surface for places from the third case. Density surface-based representations are becoming increasing pop- ular to be applied for places with indeterminate locations or bound- aries (e.g., Gao et al., 2017a; Jones et al., 2008a), which are typically derived from collective location data (e.g., geotagged images or texts from social media). In comparison, this thesis shows how the loca- tions of places can be derived as density surfaces based on collective spatial relationship knowledge with search space models. The ma- jor limitation of this georeferencing approach is the relatively small training dataset for deriving the contextualized probabilistic search spaces, particularly under certain context criteria. Although the pre- sented models are designed to be generalized enough and only re- quire a small number of training samples, a richer training dataset is still expected to further increase the georeferencing performance.
8.1.4 Place Knowledge Querying
As a knowledge base with collective information about place that people found worth describing, a PGD should also support querying of the stored knowledge for utilization. Therefore, place knowledge query is selected as the last task of this thesis. The corresponding subtasks include identifying types of queries that can be answered by the knowledge captured in the system, as well as developing query algorithms. Another challenge is querying by context, since the cap- tured information in a PGD may originally be from different conversa- tional contexts, it may have different degrees of relevance regarding a query. Conventional queries by exact-matching will return results aggregated from possibly multiple source contexts, and some of such results may be less useful. Consequently, it is necessary to consider contextualized query mechanisms.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 8.2 summary of contributions and evaluation against hypothesis 153
Chapter 6 identifies types of queries that can be answered by the knowledge captured in a PGD, explains two complementary methods for querying a PGD, as well as develops a contextualized query algo- rithm that can be applied to the identified types of queries. The query framework is able to translate structured queries into graph traversal queries that matches nodes, edges, as well as paths by their types and properties regarding the query criteria. Query examples are pro- vided for demonstration of the query framework. The query contextu- alization approach relies on the captured contextual information in a PGD for filtering the stored knowledge according to contexts. Given a query, it first retrieves candidate results (i.e., nodes, edges, paths, and properties) from the database through graph traversal. Then, it ana- lyzes these candidates and groups them by contexts. Finally, a dialog- based query approach is developed for letting users choose from the identified contexts and return results accordingly. The contextualized query approach is evaluated by comparison to query results that are delivered without the dialog-based approach that returns results that are not contextualized. The implemented query framework returns results that can be rep- resented by tabular attributes and values, which can be exported to other GISs, spatial databases, and spatial applications. The knowledge captured in a PGD can then be used to facilitate a wide range of appli- cations through querying. Again, the ultimate goal of a place-based information system is a spatially intelligent machine that is capable of understanding and communicating about place knowledge smoothly with a human. Regarding a PGD as the knowledge base supporting such an intelligent machine, while the previous three tasks focus on information modelling and processing, which can be regarded as un- derstanding, querying is one of the core parts of communicating. Finally, the proposed query framework in this chapter, particularly the match- ing mechanism, could be further improved in future work. Currently only exact matching of the specified criteria in a query is considered, while in the future additional mechanisms can be developed for the purpose of query expansion.
8.2 summary of contributions and evaluation against hypothesis
This thesis has addressed the four identified major challenges in mod- elling, reasoning with, georeferencing, and querying human place knowledge extracted from NL place descriptions. The major contri- butions and outcomes are summarized as follows:
• A designed PGD model for the efficient storage and query of place knowledge extracted from place descriptions. The supe- riority of the model has been demonstrated by comparing its performances to a competitive model in reasoning, georeferenc- ing, and querying tasks; • A spatial relational reasoning framework for maintaining rela- tional consistency in transactions happened within a PGD. The
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 154 discussion and conclusions
framework is defined over four spatial relation families lever- aging relational composition, path knowledge, and consistency rules.
• A multi-step georeferencing approach for locating all the stored places in a PGD, regardless of whether they are referred by gazetteered references or not, with sub-contributions including: a novel clustering algorithm for place disambiguation; formal and probabilistic-based spatial relationship search space mod- els; a weighted multi-value gazetteer matching approach; and an approach to derive a density surface to represent the approx- imate location of a place.
• An analysis of queries about place knowledge, a query frame- work for different types of queries based on two complementary query approaches, as well as a dialog-based querying interface for filtering query result regarding contexts; and
• An implemented PGD management system that integrates the above functionalities into a processing chain, together with sev- eral other functionalities such as database visualization and place mapping.
In the remaining part of this section, the contributions are com- pared and evaluated against the hypotheses identified in Chapter 1. Discussions are followed by the listed hypotheses below.
•A PGD model can be developed for capturing place-related and contextual information extracted from place descriptions, and the model could overcome limitations of the previously- proposed PG model and achieve better results in reasoning, geo- referencing, and querying tasks.
• Spatial relational consistency can be preserved after transac- tions (i.e., creation, update, deletion) happen within a PGD through relational inference and consistency reasoning, and the place database will always end up in a consistent state.
• The PGD model allows achieving higher precision and re- call than typical TR approaches when georeferencing places from collective place descriptions, through considering non- gazetteered references and semantics of spatial relationships.
• The PGD model allows answering different types of queries including ones that cannot be answered by the previously- proposed PG model. Moreover, query results can be contextu- alized based on the captured context information.
For the first hypothesis, a restructured and extended PG model is proposed in Chapter 3. The model additionally captures eight types of information that have been identified from place descrip- tions, which are not captured in the basic model. These types of in- formation can be classified as place-related information (e.g., place
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 8.2 summary of contributions and evaluation against hypothesis 155 semantics, affordances, and characteristics), contextual information (e.g., places and spatial relationships from the same discourses, de- scription theme), and others (e.g., routes, non-binary spatial relation- ships). The advantages of the model has been demonstrated in Chap- ters 4, 5, and 6 for reasoning, georeferencing, and querying tasks. Specifically, the extended model can be used to derive more con- strained locations of places for georeferencing; it captures additional information such as reference frames to be used in maintaining rela- tional consistency reasoning; and it is capable of answering additional spatial queries that cannot be answered by the basic PG model. There- fore, the superiority of the extended model has been confirmed for the three tasks, and the first hypothesis is proven. For the second hypothesis, a reasoning framework for spatial re- lationship inference as well as consistency checking for four types of spatial relations (cardinal direction, relative direction, qualitative distance, and topological) is developed in Chapter 4. The chapter first presents composition rules for each single spatial relation family, and thus enables inferring new spatial relationship based on exist- ing relational knowledge over paths. Then, a path-based consistency checking algorithm is presented. In addition, the chapter also extends the framework to support cross family reasoning, and the purpose is to test and compare the results to single family cases. According to the experiments, the resulting system is reliable in identifying and flagging any local inconsistent relationships within its pragmatically limited query depth during transactions, and false positives were ac- tually due to a more flexible use of language than this system was committing to. Global consistency, however, cannot be ensured due to the computational complexity. Nevertheless, human spatial knowl- edge are typically local, and experiments show that after just a few steps of composition, the inferred relationships quickly end up with the universal set anyway, and thus the urgency to test consistency over longer paths is reduced. The experiments have demonstrated the feasibility of reasoning and identifying relational contradictions with human qualitative spatial relationship knowledge using a graph database. In the future, the model can be refined and extended with research on contextualized NL spatial relationship interpretation and modelling, which currently remains a significant challenge. For the third hypothesis, a three-step comprehensive georeferenc- ing approach is presented in Chapter 5. The approach is designed for georeferencing all places in a PGD from the three identified situations. Therefore, the approach overcomes the limitation of conventional TR approaches presented in the literature, which typically only consider gazetteered place names that are coarser than a certain level of spa- tial granularity. The advantages of the approach are from two aspects: for types of places that are considered in conventional TR studies, the presented approach in this thesis is able to achieve higher precision and recall since it leverages the spatial relationship knowledge among places for refinement; and the approach is able to georeference other types of places (i.e., gazetteered places not referred by gazetteered names, or non-gazetteered places) and therefore further increases the
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 156 discussion and conclusions
recall of place georeferencing. The first advantage is demonstrated by the first step of the georeferencing approach (a novel density-based clustering algorithm), as well as the second and third step (using spa- tial relationship for further disambiguation). The second advantage is demonstrated by showing that places not referred by gazetteered names can also be georeferenced with a reasonable level of precision and distance error. Therefore, the third hypothesis is proven. For the last hypothesis, place queries in a PGD have been iden- tified and categorized, and a contextualized querying approach is introduced in Chapter 6. The classified queries are about querying node knowledge, edge knowledge, path knowledge (i.e., connection between different nodes), property value, or their combinations. The developed querying framework supports all these types of queries, and it further integrates the dialog-based interface in order to allow filtering and contextualizing query results according to user input. Demonstrations of the query framework is given using several query examples. When comparing querying using the proposed PGD model and the basic PG model, only part of the identified query types are answerable using the basic PG model. This is because some relevant information required by these queries, including contextual informa- tion, is either ignored or lost during the modelling stage of the ba- sic PG model. For example, relative direction relations have lost their reference frames and directions, and the number of times places or spatial relationships are mentioned by people is not stored using the basic PG model. Therefore, the last hypothesis is proven.
8.3 limitations and future work
This section is divided into three parts. The first two subsections iden- tify the two major limitations of this thesis, as well as future research directions regarding these limitations. The third subsection discusses several promising future research topics that built on the approaches and results presented in this thesis.
8.3.1 Natural Language Processing for Place Knowledge Extraction
Although spatial language has been comprehensively studied in linguistic and spatial cognition communities, it is relatively under- researched in the field of natural language processing (NLP). Despite the fact that spatial language processing is beyond the scope of this thesis, it is critical for automating the PGD construction process intro- duced in Chapter 3. Currently, the process of knowledge extraction from place descriptions using parsers is still limited in terms of both accuracy as well as the type of information that can be extracted, and thus the extracted data requires manual intervention in order to be used for the experiments in this thesis. Without a fully automated and accurate extraction process the proposed PGD model is difficult to be tested and evaluated against other real-world datasets. Specifi-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 8.3 limitations and future work 157 cally, two limitations are identified below, which may be overcome by future research. On one hand, the performance of place reference and spatial re- lationship extraction is expected to be further improved by employ- ing the state-of-the-art machine learning techniques. Machine learn- ing seems suitable for spatial language processing, and it has been leveraged already by the parser used in this thesis. For instance, it is promising to apply deep learning, which has pushed the record of results for many machine learning problems, with neural network models that were used for information extraction (IE) and part-of- speech (POS) tagging problems before, such as recurrent neural net- works (RNN) or long short-term memory (LSTM). For instance, it is ob- served that the current parser used in this thesis is limited for extract- ing non-binary spatial relationships. There is also no existing parser to automatically map the extracted spatial relationship expressions in NL to the formal spatial relationships considered in this thesis, and for experimental purposes this is done by a manually implemented dictionary in this thesis. On the other hand, several types of information identified in Sec- tion 3.2.1 cannot be automatically extracted yet due to the lack of parsers. Therefore, parsers developed in future research can improve the process of automatically constructing a PGD from raw place de- scriptions. Examples of such types of information are reference direc- tions, routes, place conceptualizations, and human activities associ- ated with places. Machine learning techniques being able to perform IE and POS tagging may be able to address these problems, such as hidden Markov model (HMM) or conditional random fields (CRF). For example, it may be possible to extract motion verbs for identifying routes, or to extract motion phrases for identifying human activities at places.
8.3.2 Contextualized Spatial Relationship Modelling
Interpreting and modelling spatial relationships from NL expressions is one of the core challenges of this thesis, both for the reasoning task in Chapter 4 and for the georeferencing task in Chapter 5. It also re- mains an open challenge in relevant research fields and applications, such as QSR in the field of Artificial Intelligence, geographic informa- tion retrieval (GIR), and location-based query engines. For spatial reasoning from NL, future work requires an even deeper consideration of context. While consistency rules can be set out from a logical perspective, such rules do not always reflect pragmatics from a linguistic perspective. The usage as well as interpretations of spa- tial relationships in NL are flexible and context-dependent. Examples have already been given in Section 4.4. Therefore, it is important to study what the contextual factors are and how they affect the mean- ing of NL spatial relationship expressions. Reasoning models based on other types of logic, such as fuzzy, probabilistic, or defeasible logic-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 158 discussion and conclusions
based ones may also be employed to better model the semantics of spatial relationships from NL expressions. For the task of place georeferencing, the major limitation of the proposed approach is the relatively small training dataset for deriv- ing the contextualized probabilistic search spaces, particularly under certain context criteria. Also, only contextual factors that are auto- matically obtainable from an input PGD are considered, while ones identified in the literature that could also affect the semantics of spa- tial relationships are not. Therefore, the contextualized probabilistic search space models proposed in this thesis may be further improved by considering these factors in the future, if the values of these fac- tors can be extracted from place descriptions or their metadata. In addition, the proposed contextualized search spaces are derived in a manner that no link has been established between search spaces for the same spatial relationship from different contexts. In the future, methods could be developed to set up such link through studying how each contextual factor, or a combination of contextual factors, affects the search spaces of spatial relationships.
8.3.3 Application of a place graph database
Regarding a PGD as a human knowledge base about place, the next step is to explore how it can be used for facilitating place-related research and services. This subsection discusses four research direc- tions.
route description generation A PGD is extracted from col- lections of descriptions about some environments, possibly route de- scription as well. Therefore, it may be used to generate NL route di- rections for these environments using the place references and their spatial relationships stored. Instead of simply selecting the origin and destination places in the database and find the shortest path, ad- ditional challenges exist that should be considered as well, such as the identification and choice of landmarks to be included in the de- scriptions (Nothegger et al., 2009; Raubal and Winter, 2002), switch of spatial granularity (Tomko and Winter, 2006), NL generation, and balancing between descriptions complexity and the amount of infor- mation for better understandability. The ability to generate NL route descriptions is helpful for assisting and improving current applica- tions such as navigational services. Some necessary information for these identified challenges is already captured in a PGD in order to generate such a description. For instance, the number of times places are mentioned in certain contexts as well as the number of spatial re- lationships associated with them can be used for the identification of local landmarks. A landmark identification methodology has already been proposed for the basic PG model (Kim, Vasardani, and Winter, 2017a).
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 8.3 limitations and future work 159 authoritative dataset enrichment A PGD can be lever- aged to enrich authoritative datasets, such as gazetteers and address databases, with people’s local geographic knowledge. Such knowl- edge is helpful for various research and applications that rely on these datasets, such as web searching and GIR. Another application is the automated location of callers to emergency authorities during accidents or a crisis, which can quickly fail when facing vernacular place descriptions with non-gazetteered place references and quali- tative spatial relationships. The standard available geographic infor- mation systems (such as national address files) used in such situa- tions are possibly not detailed enough for localization with regard to vernacular or granularity. In addition, location-based search engines are currently forced to interpret place-related NL queries as points or crisp polygons, and ignore qualitative spatial relationships. The geo- referencing and querying approaches developed in this thesis provide insights for better interpretations of these queries, and derive more appropriate place information. mapping and analyzing place knowledge A feasible ap- proach to georeference places in a PG is presented in this thesis. Other than being beneficial to existing location-based research and services discussed above, it also opens opportunities for mapping and analy- sis of human place knowledge. It also allows visualizing such knowl- edge, for example, as probabilistic-based density surfaces. Examples include mapping places that are frequently mentioned in certain con- texts, or human activities at places. Since a PGD provides a natural structure for capturing qualitative spatial information among places, it also provides unique opportunities for both quantitative analysis (through georeferencing) as well as qualitative analysis, regarding the emerging concept of qualitative GIS (Garnett and Kanaroglou, 2016). Another example of mapping human place knowledge is emer- gency content mapping through user-generated content. Existing ap- proaches for this task often rely on analyzing social media texts and are typically based on user geotagging information combined with keyword analysis. A system that is able to interpret NL spatial expres- sions could be useful for this task, for instance in situations when user geotagging information is unavailable. Therefore, this thesis of- fers methodologies and insights for future research in the field of emergency content mapping. A PGD itself can be used as a model for analyzing and mapping collective descriptions of emergency content as well. gis with natural language input This thesis contributes to better understanding of NL place descriptions as input to GIS in gen- eral. The developed PGD model captures human place knowledge and can be used to build NL interfaces to communicate the locations of places. Driven by the ubiquity of spatial information in mobile and web services, a plethora of applications will benefit from com- puters being able to understand human place knowledge in human- computer interactions, such as web search, communication in emer-
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 160 discussion and conclusions
gency response, or taking to a robot or an autonomous car. Such human knowledge is currently unavailable or limited in these ap- plications. These applications typically struggle with unstructured inputs, vernacular references to places, qualitative spatial relations, and contexts, while this thesis provides solutions as well as insights to these problems. For example, a PGD supporting an autonomous ve- hicle can be used to establish links between passengers’ identities and the place each one of them calls my office, and consequently the same place reference is infused with contextualized meanings. Compared to existing applications that pre-defines place semantics for users in a top-down manner (e.g., a user is able to link locations to several pre- defined places such as office or home), the PGD model proposed in this thesis contextualize the semantics of place in a bottom-up man- ner. A PGD collects human knowledge about place from NL descrip- tions without any pre-defined constrains, while it allows for future tracking of such knowledge in order to use the knowledge for appli- cations or analysis. Therefore, this thesis will also feed into spatially intelligent systems with NL interfaces.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] BIBLIOGRAPHY
Adams B. and K. Janowicz (2012). “On the Geo-Indicativeness of Non- Georeferenced Text.” In: Proceedings of the Sixth International Con- ference on Weblogs and Social Media. Ed. by J. G. Breslin et al. The AAAI Press, pp. 375–378. Adams B. and K. Janowicz (2015). “Thematic signatures for cleansing and enriching place-related linked data.” In: International Journal of Geographical Information Science 29.4, pp. 556–579. Adams B. and G. McKenzie (2013). “Inferring thematic places from spatially referenced natural language descriptions.” In: Crowd- sourcing Geographic Knowledge: Volunteered Geographic Information (VGI) in Theory and Practice. Ed. by D. Z. Sui, S. Elwood, and M. F. Goodchild. Springer, pp. 201–221. Adelfio M. D. and H. Samet (2013a). “GeoWhiz: Toponym resolu- tion using common categories.” In: Proceedings of the 21st ACM SIGSPATIAL International Conference on Advances in Geographic In- formation Systems - SIGSPATIAL’13. ACM, pp. 532–535. Adelfio M. D. and H. Samet (2013b). “Structured toponym resolution using combined hierarchical place categories.” In: Proceedings of the 7th Workshop on Geographic Information Retrieval - GIR’13. ACM, pp. 49–56. Agarwal P. (2005). “Operationalising‘Sense of Place’as a Cognitive Operator for Semantics in Place-Based Ontologies.” In: Proceed- ings of the International Conference of Spatial Information Theory - COSIT’05. Ed. by A. G. Cohn and D. M. Mark. Vol. 3693. Springer, pp. 96–114. Agnew J. (2011). “Space and place.” In: Handbook of Geographical Knowledge 2011, pp. 316–331. Alazzawi A. N., A. I. Abdelmoty, and C. B. Jones (2012). “What can I do there? Towards the automatic discovery of place-related ser- vices and activities.” In: International Journal of Geographical Infor- mation Science 26.2, pp. 345–364. Amitay E. et al. (2004). “Web-a-Where: Geotagging Web Content.” In: Proceedings of the 27th Annual International ACM SIGIR Con- ference on Research and Development in Information Retrieval. ACM, pp. 273–280. Anastácio I., B. Martins, and P. Calado (2009). “Classifying documents according to locational relevance.” In: Progress in Artificial Intelli- gence, 14th Portuguese Conference on Artificial Intelligence - EPIC’09. Vol. 5816. Springer, pp. 598–609.
161
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 162 bibliography
Angiulli F. (2006). “Clustering by exceptions.” In: Proceedings of the 21th National Conference on Artificial Intelligence. Vol. 21. 1. Menlo Park, CA; Cambridge, MA; London; AAAI Press; MIT Press; 1999, pp. 312–317. Angles R. and C. Gutierrez (2008). “Survey of graph database mod- els.” In: ACM Computing Surveys 40.1, pp. 1–39. Ankerst M et al. (1999). “OPTICS: Ordering Points To Identify the Clustering Structure.” In: Proceedings of the ACM SIGMOD’99 In- ternational Conference on Management of Data. Vol. 28. 2. ACM, pp. 49–60. Ballatore A., M. Bertolotto, and D. C. Wilson (2013). “Grounding linked open data in WordNet: The case of the OSM semantic network.” In: Web and Wireless Geographical Information Systems - W2GIS’13. Ed. by S. H. L. Liang, X. Wang, and C. Claramunt. Vol. 7820. Lecture Notes in Computer Science. Springer, pp. 1– 15. Basiri A., P. Amirian, and A. Winstanley (2014). “Use of graph databases in tourist navigation application.” In: Proceedings of the 14th International Conference on Computational Science and Its Appli- cations - ICCSA’14. Ed. by B. Murgante et al. Vol. 8583. 5. Springer, pp. 663–677. Bateman J. A. et al. (2010). “A linguistic ontology of space for natural language processing.” In: Artificial Intelligence 174.14, pp. 1027– 1071. Belouaer L., D. Brosset, and C. Claramunt (2013). “Modeling spatial knowledge from verbal descriptions.” In: Proceedings of the 11th In- ternational Conference on Spatial Information Theory - COSIT’13. Ed. by T. Tenbrink et al. Vol. 8116. Lecture Notes in Computer Science. Cham, Switzerland: Springer International Publishing, pp. 338– 357. Belouaer L., D. Brosset, and C. Claramunt (2016). “From verbal route descriptions to sketch maps in natural environments.” In: Pro- ceedings of the 24th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems. Ed. by S. Ravada et al. ACM, 13:1–13:10. Bennett B. and P. Agarwal (2007). “Semantic Categories Underlying the Meaning of ‘Place’.” In: Proceedings of the 8th International Con- ference on Spatial Information Theory - COSIT’07. Ed. by S. Win- ter et al. Vol. 4736. Lecture Notes in Computer Science. Springer, pp. 78–95. Berkhin P (2006). “A survey of clustering data mining techniques.” In: Grouping Multidimensional Data. Ed. by N. C. Kogan J. and T. M. Springer, pp. 25–71. Berners-Lee T., J. Hendler, and O. Lassila (2001). “The Semantic Web.” In: Scientific American 284.5, pp. 34–43.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] bibliography 163
Bittner T. and J. G. Stell (2000). “Rough sets in approximate spa- tial reasoning.” In: Rough Sets and Current Trends in Computing - RSCTC’00. Ed. by W. Ziarko and Y. Y. Yao. Vol. 2005. Springer, pp. 445–453. Bizer C., T. Heath, and T. Berners-Lee (2011). “Linked data: The story so far.” In: Semantic Services, Interoperability and Web Applications: Emerging Concepts. IGI Global, pp. 205–227. Blei D. M., A. Y. Ng, and M. I. Jordan (2003). “Latent dirichlet alloca- tion.” In: Journal of machine Learning research 3.Jan, pp. 993–1022. Bollacker K. et al. (2008). “Freebase: A Collaboratively Created Graph Database For Structuring Human Knowledge.” In: Proceedings of the 2008 ACM SIGMOD International Conference on Management of Data. AcM, pp. 1247–1250. Brosset D., C. Claramunt, and E. Saux (2007). “A Location and Action- based Model for Route Descriptions.” In: Proceedings of the 2nd In- ternational Conference on GeoSpatial Semantics. Ed. by F. T. Fonseca, M. A. Rodríguez, and S. Levashkin. Springer, pp. 146–159. Buscaldi D. (2011). “Approaches to disambiguating toponyms.” In: SIGSPATIAL Special 3.2, pp. 16–19. Buscaldi D. and B. Magnini (2010). “Grounding toponyms in an Ital- ian local news corpus.” In: Proceedings of the 6th Workshop on Geographic Information Retrieval - GIR’10. Ed. by R. Purves, P. D. Clough, and C. B. Jones. ACM, pp. 1–15. Buscaldi D. and P. Rosso (2008a). “Map-based vs. knowledge-based toponym disambiguation.” In: Proceedings of the 2nd International Workshop on Geographic Information Retrieval - GIR ’08. Ed. by C. Jones and R. Purves. ACM, p. 19. Buscaldi D. and P. Rosso (2008b). “A conceptual density-based ap- proach for the disambiguation of toponyms.” In: International Journal of Geographical Information Science 22.3, pp. 301–313. Cai G. (2007). “Contextualization of geospatial database semantics for human-GIS interaction.” In: GeoInformatica 11.2, pp. 217–237. Campello R. J. G. B., D. Moulavi, and J. Sander (2013). “Density-based clustering based on hierarchical density estimates.” In: Advances in Knowledge Discovery and Data Mining - PAKDD’13. Ed. by J. Pei et al. Springer, pp. 160–172. Canter D. V. (1977). The psychology of place. St. Martin Press. Celeux G. and G. Govaert (1992). “A Classification EM Algorithm for Clustering and Two Stochastic Versions.” In: Computational Statistics & Data Analysis 14.3, pp. 315–332. Chen H., M. Vasardani, and S. Winter (2015). “Maintaining relational consistency in a graph-based place database.” In: CEUR Workshop Proceedings - Locate15. Vol. 1323, pp. 1–12.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 164 bibliography
Chen H., M. Vasardani, and S. Winter (2018a). “Clustering-based dis- ambiguation of fine-grained place names from descriptions.” In: GeoInformatica. Chen H., M. Vasardani, and S. Winter (2018b). “Georeferencing places from collective human descriptions using place graphs.” In: Jour- nal of Spatial Information Science 2018.17, pp. 31–62. Chen H. et al. (2018). “A Graph Database Model for Knowledge Ex- tracted from Place Descriptions.” In: ISPRS International Journal of Geo-Information 7.6, p. 221. Cheng Z, J Caverlee, and K Lee (2010). “You are where you tweet: A content-based approach to geo-locating Twitter users.” In: Pro- ceedings of the 19th ACM International Conference on Information and Knowledge Management. ACM, pp. 759–768. Cheng Z., J. Caverlee, and K. Lee (2013). “A content-driven frame- work for geolocating microblog users.” In: ACM Transactions on Intelligent Systems and Technology 4.1, pp. 1–27. Chown E., S. Kaplan, and D. Kortenkamp (1995). “Prototypes, loca- tion, and associative networks (PLAN): Towards a unified theory of cognitive mapping.” In: Cognitive Science 19.1, pp. 1–51. Clementini E and P. D. Felice (1996). “An algebraic model for spa- tial objects with indeterminate boundaries.” In: Geographic Objects With Indeterminate Boundaries 2. Ed. by P. A. Burrough and A. U. Frank, pp. 155–169. Cohn A. G. and S. M. Hazarika (2001). “Qualitative Spatial Represen- tation and Reasoning: An Overview.” In: Fundamenta Informaticae 46.1/2, p. 1. Cohn A. G. and N. M. Gotts (1996). “The ’Egg-Yolk’ Representation of Regions with Indeterminate Boundaries.” In: Geographic Objects With Indeterminate Boundaries 2. Ed. by P. A. Burrough and A. U. Frank, pp. 171–188. Cohn A. G. and J. Renz (2008). “Qualitative spatial representation and reasoning.” In: Foundations of Artificial Intelligence 3, pp. 551–596. Cohn A. G. et al. (2014). “Reasoning about Topological and Cardinal Direction Relations Between 2-Dimensional Spatial Objects.” In: Journal of Artificial Intelligence Research 51, pp. 493–532. Couclelis H et al. (1987). “Exploring the anchor-point hypothesis of spatial cognition.” In: Journal of Environmental Psychology 7.2, pp. 99–122. Couclelis H. (1992). “Location, place, region, and space.” In: Geogra- phy’s Inner Worlds 2, pp. 15–233. Cowie J. and W. Lehnert (1996). “Information extraction.” In: Commu- nications of the ACM 39.1, pp. 80–91.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] bibliography 165
Coyne B., R. Sproat, and J. Hirschberg (2010). “Spatial relations in text- to-scene conversion.” In: Computational Models of Spatial Language Interpretation, Workshop at Spatial Cognition. Cresswell T. (2014). Place: An Introduction. John Wiley & Sons, p. 232. Culotta A. and J. Sorensen (2004). “Dependency tree kernels for re- lation extraction.” In: Proceedings of the 42nd Annual Meeting on Association for Computational Linguistics - ACL’04. Association for Computational Linguistics, p. 423. Dalton J., L. Dietz, and J. Allan (2014). “Entity query feature expan- sion using knowledge base links.” In: Proceedings of the 37th Inter- national ACM SIGIR Conference on Research & Development in Infor- mation Retrieval - SIGIR’14. Ed. by S. Geva et al. ACM, pp. 365– 374. Damerau F. J. (1964). “A technique for computer detection and cor- rection of spelling errors.” In: Communications of the ACM 7.3, pp. 171–176. Davies C. et al. (2008). “User Needs and the Implications for Mod- elling Place.” In: Proceedings of the International Workshop on Com- putational Models of Place 8.3. Ed. by S. Winter, W. Kuhn, and Krüger Antonio, pp. 1–14. Delboni T. M. et al. (2007). “Semantic expansion of geographic web queries based on natural language positioning expressions.” In: Transactions in GIS 11.3, pp. 377–397. Delozier G., J. Baldridge, and L. London (2015). “Gazetteer- Independent Toponym Resolution Using Geographic Word Pro- files.” In: Proceedings of the 29th AAAI Conference on Artificial Intel- ligence. AAAI Press, pp. 2382–2388. Derungs C., D. Palacio, and R. S. Purves (2012). “Resolving fine gran- ularity toponyms : Evaluation of a disambiguation approach.” In: Proceedings of the 7th International Conference on Geographic Informa- tion Science, pp. 1–5. Derungs C. and R. S. Purves (2014). “From text to landscape: Locating, identifying and mapping the use of landscape features in a Swiss Alpine corpus.” In: International Journal of Geographical Information Science 28.6, pp. 1272–1293. Derungs C. and R. S. Purves (2016). “Mining nearness relations from an n-grams Web corpus in geographical space.” In: Spatial Cogni- tion and Computation 16.4, pp. 301–322. Duckham M. and M. Worboys (2001). “Computational structure in three-valued nearness relations.” In: Spatial Information Theory: Foundations of Geographic Information Science - COSIT’01. Ed. by D. R. Montello. Springer, pp. 76–91.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 166 bibliography
Düntsch I. (2005). “Relation algebras and their application in tem- poral and spatial reasoning.” In: Artificial Intelligence Review 23.4, pp. 315–357. Egenhofer M. J. (1991). “Reasoning about binary topological re- lations.” In: Advances in Spatial Databases - SSD’91. Springer, pp. 143–160. Egenhofer M. J. (1994). “Deriving the Composition of Binary Topolog- ical Relations.” In: Journal of Visual Languages and Computing 5.2, pp. 133–149. Egenhofer M. J. (2002). “Toward the semantic geospatial web.” In: Proceedings of the 10th ACM International Symposium on Advances in Geographic Information Systems - GIS ’02. Ed. by A. Voisard and S.-C. Chen. ACM, pp. 1–4. Egenhofer M. J. and R. D. Franzosa (1991). “Point-set topological spa- tial relations.” In: International Journal of Geographical Information Systems 5.2, pp. 161–174. Egenhofer M. J. and D. M. Mark (1995). “Naive geography.” In: Spatial Information Theory: A Theoretical Basis for GIS. Ed. by A. U Frank and W Kuhn. Springer, pp. 1–15. Eisenstein J. et al. (2010). “A Latent Variable Model for Geographic Lexical Variation.” In: Proceedings of the 2010 Conference on Empiri- cal Methods in Natural Language Processing - EMNLP’10. Vol. 11. 6. Association for Computational Linguistics, pp. 1277–1287. Elwood S., M. F. Goodchild, and D. Sui (2013). “Prospects for VGI research and the emerging fourth paradigm.” In: Crowdsourcing Geographic Knowledge: Volunteered Geographic Information (VGI) in Theory and Practice. Ed. by D. Z. Sui, S. Elwood, and M. F. Good- child. Springer, pp. 361–375. Ertöz L., M. Steinbach, and V. Kumar (2003). “Finding Clusters of Dif- ferent Sizes, Shapes, and Densities in Noisy, High Dimensional Data.” In: Proceedings of the 2003 SIAM International Conference on Data Mining. Ed. by D. Barbará and C. Kamath. SIAM, pp. 47–58. Ester, M., Kriegel, H. P., Sander, J., & Xu X (1996). “A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise.” In: Proceedings of 2nd International Conference on Knowledge Discovery and Data Mining. Ed. by E. Simoudis, J. Han, and U. Fayyad. Vol. 96. 34. AAAI Press, pp. 226–231. Fisher P. F. (1996). “Boolean and fuzzy regions.” In: Geographic Objects With Indeterminate Boundaries 2. Ed. by P. A. Burrough and A. U. Frank, pp. 87–94. Fogliaroni P. and H. Hobel (2015). “Implementing Naïve Geography via Qualitative Spatial Relation Queries.” In: Proceedings of the 18th AGILE Conference on Geographic Information Science, pp. 9–12.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] bibliography 167
Fonseca F. (2008). “Geospatial semantic web.” In: Encyclopedia of {GIS}. Ed. by S. Shekhar, H. Xiong, and X. Zhou. Springer, pp. 388–391. Frank A. U. (1992). “Qualitative spatial reasoning about distances and directions in geographic space.” In: Journal of Visual Languages & Computing 3.4, pp. 343–371. Frank A. U. (1996). “Qualitative spatial reasoning: Cardinal directions as an example.” In: International Journal of Geographical Information Systems 10.3, pp. 269–290. Frank A. U. (2003). “Ontology for spatio-temporal databases.” In: Spatio-Temporal Databases. Ed. by M. Koubarakis et al. Springer, pp. 9–77. Freksa C. (1991). “Qualitative spatial reasoning.” In: Cognitive and Lin- guistic Aspects of Geographic Space. Ed. by M. Raubal, D. M. Mark, and A. U. Frank. Springer, pp. 361–372. Freksa C. (1992). “Using orientation information for qualitative spa- tial reasoning.” In: Theories and Methods of Spatio-Temporal Reason- ing in Geographic Space. Ed. by A. U. Frank, I. Campari, and U. For- mentini. Vol. 639. Lecture Notes in Computer Science. Springer, pp. 162–178. Freksa C., J. van de Ven, and D. Wolter (2018). “Formal representation of qualitative direction.” In: International Journal of Geographical Information Science 32.12, pp. 2514–2534. Fu G., C. B. Jones, and A. I. Abdelmoty (2005a). “Building a Geo- graphical Ontology for Intelligent Spatial Search on the Web.” In: Proceedings of the International Conference on Databases and Applica- tions. Ed. by M. H. Hamza, pp. 167–172. Fu G., C. B. Jones, and A. I. Abdelmoty (2005b). “Ontology-based spatial query expansion in information retrieval.” In: Proceedings of the OTM Confederated International Conferences. Ed. by R. Meers- man and Z. Tari. Vol. 3761. Springer, pp. 1466–1482. Galton A. and J. Hood (2005). “Anchoring: A New Approach to Han- dling Indeterminate Location in GIS.” In: Proceedings of the Inter- national Conference of Spatial Information Theory - COSIT’05. Ed. by A. Cohn and D. Mark. Vol. 3693. Lecture Notes in Computer Sci- ence. Springer, pp. 1–13. Gan Q. et al. (2008). “Analysis of geographic queries in a search en- gine log.” In: Proceedings of the 1st International Workshop on Loca- tion and the Web - LOCWEB’08. ACM, pp. 49–56. Gao S. et al. (2017a). “A data-synthesis-driven method for detecting and extracting vague cognitive regions.” In: International Journal of Geographical Information Science 31.6, pp. 1245–1271. Gao S. et al. (2017b). “Constructing gazetteers from volunteered big geo-data based on Hadoop.” In: Computers, Environment and Ur- ban Systems 61, pp. 172–186.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 168 bibliography
Garbin E. and I. Mani (2005). “Disambiguating toponyms in news.” In: Proceedings of the Conference on Human Language Technology and Empirical Methods in Natural Language Processing - HLT ’05. Asso- ciation for Computational Linguistics, pp. 363–370. Gärdenfors P. (2014). The geometry of meaning: Semantics based on con- ceptual spaces. MIT Press. Garnett R. and P. Kanaroglou (2016). “Qualitative GIS: An open framework using spatialite and open source GIS.” In: Transactions in GIS 20.1, pp. 144–159. Geeraerts D. and H. Cuyckens (2007). The Oxford handbook of cognitive linguistics. Oxford University Press. Geiß J. et al. (2015). “The Wikipedia location network: overcoming borders and oceans.” In: Proceedings of the 9th Workshop on Geo- graphic Information Retrieval - GIR’15. Ed. by R. S. Purves and C. B. Jones. ACM, pp. 1–3. Gerevini A. and J. Renz (1998). “Combining topological and qualita- tive size constraints for spatial reasoning.” In: Proceedings of the International Conference on Principles and Practice of Constraint Pro- gramming. Ed. by M. J. Maher and J.-F. Puget. Springer, pp. 220– 234. Gerevini A. and J. Renz (2002). “Combining topological and size in- formation for spatial reasoning.” In: Artificial Intelligence 137.1-2, pp. 1–42. Gibson J. J. (1979). The ecological approach to visual perception. Houghton Mifflin. Golledge R. G. (1997). Spatial behavior: A geographic perspective. Guil- ford Press. Goodchild M. F. and L. L. Hill (2008). “Introduction to digital gazetteer research.” In: International Journal of Geographical Infor- mation Science 22.10, pp. 1039–1044. Goodchild M. F. (2007). “Citizens as sensors: The world of volun- teered geography.” In: GeoJournal 69.4, pp. 211–221. Goodchild M. F. (2011). “Formalizing Place in Geographic Infor- mation Systems.” In: Communities, Neighborhoods, and Health. Springer, pp. 21–33. Grothe C. and J. Schaab (2009). “Automated footprint generation from geotags with kernel density estimation and support vector machines.” In: Spatial Cognition and Computation 9.3, pp. 195–211. Guha S., R. Rastogi, and K. Shim (1998). “CURE: an efficient clus- tering algorithm for large databases.” In: Proceedings of the 1998 ACM SIGMOD International Conference on Management of Data - SIGMOD’98. Vol. 27. 2, pp. 73–84.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] bibliography 169
Guo Q, Y Liu, and J Wieczorek (2008). “Georeferencing locality de- scriptions and computing associated uncertainty using a proba- bilistic approach.” In: International Journal of Geographical Informa- tion Science 22.10, pp. 1067–1090. Güting R. H. (1994). “GraphDB: Modeling and querying graphs in databases.” In: Proceedings of 20th International Conference on Very Large Data Bases. Vol. 94, pp. 12–15. Habib M. B. and M van Keulen (2012). “Improving Toponym Disam- biguation by Iteratively Enhancing Certainty of Extraction.” In: Proceedings of the 4th International Conference on Knowledge Discov- ery and Information Retrieval - KDIR 2012. Ed. by A. L. N. Fred and J. Filipe. Barcelona, Spain: SciTePress, pp. 399–410. Hall M. M., P. D. Smart, and C. B. Jones (2011). “Interpreting spatial language in image captions.” In: Cognitive Processing 12.1, pp. 67– 94. Hamzei E. et al. (2018). “Deriving place graphs from spatial databases.” In: CEUR Workshop Proceedings - Locate18. Ed. by S. Peters and K. Khoshelham. Vol. 2087, pp. 25–32. Han B., P. Cook, and T. Baldwin (2014). “Text-Based Twitter User Ge- olocation Prediction.” In: Journal of Artificial Intelligence Research 49, pp. 451–500. Harrison S. and P. Dourish (1996). “Re-Place-ing Space : The Roles of Place and Space in Collaborative Systems.” In: Proceedings of the 1996 ACM Conference on Computer Supported Cooperative Work. Ed. by M. S. Ackerman. Vol. 7. ACM, pp. 67–76. Hartigan J. a. and M. a. Wong (1979). “Algorithm AS 136: A k-means clustering algorithm.” In: Journal of the Royal Statistical Society 28.1, pp. 100–108. Hastings J. T. (2008). “Automated conflation of digital gazetteer data.” In: International Journal of Geographical Information Science 22.10, pp. 1109–1127. Herskovits A. (1985). “Semantics and Pragmatics of Locative Expres- sions.” In: Cognitive Science 9.3, pp. 341–378. Hill L. L. (2000). “Core Elements of Digital Gazetteers: Placenames, Categories, and Footprints.” In: Research and Advanced Technology for Digital Libraries. Ed. by J. L. Borbinha and T. Baker. Vol. 1923. Lecture Notes in Computer Science. Springer, pp. 280–290. Hobel H. and P. Fogliaroni (2016). “Extracting semantics of places from user generated content.” In: Proceedings of the 19th AGILE International Conference on Geographic Information Science. Hollenstein L. and R. Purves (2010). “Exploring place through user- generated content: Using Flickr to describe city cores.” In: Journal of Spatial Information Science 2010.1, pp. 21–48.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 170 bibliography
Hong J.-H., M. Egenhofer, and A. U. Frank (1995). “On the robust- ness of qualitative distance-and direction-reasoning.” In: Auto- carto Conference, pp. 301–310. Hua H., J. Renz, and X. Ge (2018). “Qualitative Representation and Reasoning over Direction Relations across Different Frames of Reference.” In: Proceedings of the 16th International Conference on Principles of Knowledge Representation and Reasoning - KR’18. Janowicz K and C Keßler (2008). “The role of ontology in improv- ing gazetteer interaction.” In: International Journal of Geographical Information Science 22.10, pp. 1129–1157. Janowicz K. (2006). “Sim-DL: Towards a Semantic Similarity Measure- ment Theory for the Description Logic ALCNR in Geographic Information Retrieval.” In: On the Move to Meaningful Internet Sys- tems 2006: OTM 2006 Workshops. Ed. by Z. Tari. Springer, pp. 1681– 1692. Janowicz K. et al. (2010). “Semantic enablement for spatial data in- frastructures.” In: Transactions in GIS 14.2, pp. 111–129. Jarvella R. J. and W. Klein (1982). Speech, place, and action: Studies of deixis and related topics. John Wiley & Sons. Jiang J. J. and D. W. Conrath (1997). “Semantic Similarity Based on Corpus Statistics and Lexical Taxonomy.” In: Proceedings of the 10th International Conference on Research in Computational Linguis- tics - ROCLING’97. Jones C. B. et al. (2008a). “Modelling vague places with knowledge from the Web.” In: International Journal of Geographical Information Science 22.10, pp. 1045–1065. Jones C. B. and R. S. Purves (2007). “Geographical information re- trieval.” In: International Journal of Geographical Information Science 22.3, pp. 219–228. Jones C. B. et al. (2004). “The SPIRIT Spatial Search Engine: Architec- ture, Ontologies and Spatial Indexing.” In: Proceedings of the 3rd International Conference on Geographic Information Science. Springer, pp. 125–139. Jones R. et al. (2008b). “Geographic intention and modification in web search.” In: International Journal of Geographical Information Science 22.3, pp. 229–246. Jordan T et al. (1998). “An Affordance-Based Model of Place in GIS.” In: Proceedings of the 8th International Symposium on Spatial Data Handling. Vol. 98, pp. 98–109. Karimzadeh M. et al. (2013). “GeoTxt: a web API to leverage place references in text.” In: Proceedings of the 7th Workshop on Geographic Information Retrieval - GIR’13. Ed. by C. Jones and R. Purves. ACM, pp. 72–73.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] bibliography 171
Karypis G., E.-H. Han, and V. Kumar (1999). “Chameleon: Hierar- chical clustering using dynamic modeling.” In: Computer 32.8, pp. 68–75. Keßler C., K. Janowicz, and M. Bishr (2009). “An agenda for the next generation gazetteer: Geographic information contribution and retrieval.” In: Proceedings of the 17th ACM SIGSPATIAL Interna- tional Conference on Advances in Geographic Information Systems - GIS’09. Ed. by D. Agrawal et al. ACM, pp. 91–100. Khan A., M. Vasardani, and S. Winter (2013). “Extracting Spatial In- formation From Place Descriptions.” In: Proceedings of The First ACM SIGSPATIAL International Workshop on Computational Models of Place - COMP’13. Ed. by S. Scheider et al., pp. 62–69. Kim J.-D. et al. (2009). “Overview of BioNLP’09 shared task on event extraction.” In: Proceedings of the Workshop on BioNLP Shared Task - BioNLP’09. Association for Computational Linguistics, pp. 1–9. Kim J., M. Vasardani, and S. Winter (2015). “Harvesting large corpora for generating place graphs.” In: International Workshop on Cogni- tive Engineering for Spatial Information Processes. Vol. 12. Kim J., M. Vasardani, and S. Winter (2016). “From descriptions to depictions: A dynamic sketch map drawing strategy.” In: Spatial Cognition and Computation 16.1, pp. 29–53. Kim J., M. Vasardani, and S. Winter (2017a). “Landmark Extraction from Web-Harvested Place Descriptions.” In: Künstliche Intelligenz 31.2, pp. 151–159. Kim J., M. Vasardani, and S. Winter (2017b). “Similarity matching for integrating spatial information extracted from place descrip- tions.” In: International Journal of Geographical Information Science 31.1, pp. 56–80. Klippel A. and D. R. Montello (2007). “Linguistic and Nonlinguistic Turn Direction Concepts.” In: Proceedings of the 8th International Conference on Spatial Information Theory - COSIT’07. Ed. by S. Win- ter et al. Springer, pp. 354–372. Klippel A. et al. (2004). “Direction concepts in wayfinding assistance systems.” In: Proceedings of the Workshop on Artificial Intelligence in Mobile Systems, pp. 1–8. Knauff M, R Rauh, and C Schlieder (1995). “Preferred mental models in qualitative spatial reasoning: A cognitive assessment of Allen’s calculus.” In: Proceedings of the 17th Annual Conference of the Cogni- tive Science Society, pp. 200–205. Knauff M. et al. (1998). “Mental Models in Spatial Reasoning.” In: Spatial Cognition: An Interdisciplinary Approach to Representing and Processing Spatial Knowledge. Ed. by C. Freksa, C. Habel, and K. F. Wender. Springer, pp. 267–292.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 172 bibliography
Kohonen T. (1998). “The self-organizing map.” In: Neurocomputing 21.1-3, pp. 1–6. Kordjamshidi P., M. Van Otterlo, and M.-F. Moens (2011). “Spatial role labeling: Towards extraction of spatial relations from natural language.” In: ACM Transactions on Speech and Language Processing 8.3, 4:1–4:36. Kordjamshidi P. et al. (2011). “Relational learning for spatial relation extraction from natural language.” In: Proceedings of the Interna- tional Conference on Inductive Logic Programming. Ed. by S. Mug- gleton, A. Tamaddoni-Nezhad, and F. A. Lisi. Springer, pp. 204– 220. Kuhn W (2001). “Ontologies in support of activities in geographical space.” In: International Journal of Geographical Information Science 15.7, pp. 613–631. Kuhn W. (2007). “An Image-Schematic Account of Spatial Cate- gories.” In: Proceedings of the 8th International Conference on Spatial Information Theory - COSIT’07. Springer, pp. 152–168. Kuhn W. (2012). “Core concepts of spatial information for transdisci- plinary research.” In: International Journal of Geographical Informa- tion Science 26.12, pp. 2267–2276. Kuipers B and Y. T. Byun (1991). “A robot exploration and map- ping strategy based on a semantic hierarchy of spatial representa- tions.” In: Journal of Robotics and Autonomous Systems 8.1-2, pp. 47– 63. Kuipers B., D. G. Tecuci, and B. J. Stankiewicz (2003). “The Skeleton In The Cognitive Map.” In: Environment and Behavior 35.1, pp. 81– 106. Kulik L. (2001). “A Geometric Theory of Vague Boundaries Based on Supervaluation.” In: Spatial Information Theory: Foundations of Geographic Information Science. Ed. by D. R. Montello. Springer, pp. 44–59. Laere O. V. et al. (2014). “Georeferencing Wikipedia Documents Us- ing Data from Social Media Sources.” In: ACM Transactions on Information Systems 32.3, pp. 1–32. Lakoff G. (1987). Women, fire, and dangerous things. University of Chicago press. Landau B and R Jackendoff (1993). “What and Where in Spatial Lan- guage and Spatial Cognition.” In: Behavioral and Brain Sciences 16.2, pp. 255–265. Landau B. (1994). “Where’s what and what’s where: the language of objects in space.” In: Lingua 92, pp. 259–296. Langacker R. W. (1987). Foundations of cognitive grammar: Theoretical prerequisites. Vol. 1. Stanford university press.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] bibliography 173
Leidner J. L. and M. D. Lieberman (2011). “Detecting geographical ref- erences in the form of place names and associated spatial natural language.” In: SIGSPATIAL Special 3.2, pp. 5–11. Leidner J. L. and Others (2004). “Toponym resolution in text: Which Sheffield is it?” In: Proceedings of the the 27th annual international ACM SIGIR conference - SIGIR’04. Ed. by M. Sanderson et al. ACM, p. 602. Leidner J. L., G. Sinclair, and B. Webber (2003). “Grounding spatial named entities for information extraction and question answer- ing.” In: Proceedings of the Workshop on Analysis of Geographic Ref- erences - HLT/NAACL’03. Vol. 1. Association for Computational Linguistics, pp. 31–38. Leidner J. L. (2007). Toponym resolution in text: Annotation, evalua- tion and applications of spatial grounding of place names. Universal- Publishers, p. 292. Levenshtein V. I. (1966). “Binary Codes Capable of Correcting Dele- tions, Insertions, and Reversals.” In: Soviet Physics Doklady 10.8, pp. 707–710. Levinson S. C. (1996). “Frames of reference and Molyneux’s question: Crosslinguistic evidence.” In: Language and space 109, pp. 109–169. Lieberman M. D., H. Samet, and J. Sankaranarayanan (2007). “STEW- ARD: architecture of a spatio-textual search engine.” In: Proceed- ings of the 15th Annual ACM International Symposium on Advances in Geographic Information Systems - GIS’07. Ed. by H. Samet, C. Shahabi, and M. Schneider. ACM, pp. 186–193. Ligozat G. (1998). “Reasoning about cardinal directions.” In: Journal of Visual Languages and Computing 9.1, pp. 23–44. Ligozat G. and J. Renz (2004). “What is a qualitative calculus? A gen- eral framework.” In: PRICAI: Trends in Artificial Intelligence. Ed. by C. Zhang, H. W. Guesgen, and W.-K. Yeap. Springer, pp. 53–64. Liu F., M. Vasardani, and T. Baldwin (2014). “Automatic Identification of Locative Expressions from Social Media Text: A Comparative Analysis.” In: Proceedings of the 4th International Workshop on Loca- tion and the Web - LocWeb’14. Ed. by D. Ahlers, E. Wilde, and B. Martins, pp. 9–16. Liu J. and D. Inkpen (2015). “Estimating User Location in Social Me- dia with Stacked Denoising Auto-encoders.” In: Proceedings of the 1st Workshop on Vector Space Modeling for Natural Language Process- ing. Ed. by P. Blunsom et al., pp. 201–210. Liu W. et al. (2009a). “Combining RCC-8 with Qualitative Direc- tion Calculi: Algorithms and Complexity.” In: Proceedings of the 21st International Joint Conference on Artificial Intelligence. Ed. by C. Boutilier, pp. 854–859.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 174 bibliography
Liu Y et al. (2005). “On internal cardinal direction relations.” In: Pro- ceedings of the International Conference on Spatial Information Theory - COSIT’05. Ed. by A. G. Cohn and D. M. Mark. Vol. 3693. Lecture Notes in Computer Science. Springer, pp. 283–299. Liu Y et al. (2009b). “Positioning localities based on spatial asser- tions.” In: International Journal of Geographical Information Science 23.11, pp. 1471–1501. Loureiro V., I. Anastácio, and B. Martins (2011). “Learning to resolve geographical and temporal references in text.” In: Proceedings of the 19th ACM SIGSPATIAL International Symposium on Advances in Geographic Information Systems - GIS ’11. Ed. by I. F. Cruz et al. ACM, p. 349. Lynch K. (1960). The image of the city. MIT press. Mani I. et al. (2010). “{SpatialML}: annotation scheme, resources, and evaluation.” In: Language Resources and Evaluation 44.3, pp. 263– 280. Mark D. M. and M. J. Egenhofer (1994). “Calibrating the meanings of spatial predicates from natural language: Line-region relations.” In: Proceedings of Spatial Data Handling. Vol. 1, pp. 538–553. Mark D. M. et al. (1995). “Evaluating and refining computational models of spatial relations through cross-linguistic human- subjects testing.” In: Spatial Information Theory: A Theoretical Basis for GIS - COSIT’95. Ed. by A. U. Frank and W. Kuhn. Vol. 988. Springer, pp. 553–568. McKenzie G., K. Janowicz, and B. Adams (2014). “A weighted multi- attribute method for matching user-generated Points of Interest.” In: Cartography and Geographic Information Science 41.2, pp. 125– 137. McNamara T. P. (1991). “Memory’s view of space.” In: Psychology of Learning and Motivation - Advances in Research and Theory 27.C, pp. 147–186. Melo F. and B. Martins (2017). “Automated Geocoding of Textual Doc- uments: A Survey of Current Approaches.” In: Transactions in GIS 21.1, pp. 3–38. Moncla L. et al. (2014). “Geocoding for texts with fine-grain to- ponyms: an experiment on a geoparsed hiking descriptions cor- pus.” In: Proceedings of the 22nd ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems. Ed. by Y. Huang et al., pp. 183–192. Moncla L. et al. (2016). “Reconstruction of itineraries from annotated text with an informed spanning tree algorithm.” In: International Journal of Geographical Information Science 30.6, pp. 1137–1160. Moratz R., F. Dylla, and L. Frommberger (2005). “A Relative Orienta- tion Algebra with Adjustable Granularity.” In: Proceedings of the
[ April 29, 2019 at 20:02 – classicthesis version 5 ] bibliography 175
Workshop on Agents in Real-Time and Dynamic Environments - IJ- CAI’05, pp. 61–70. Moratz R. and M. Ragni (2008). “Qualitative spatial reasoning about relative point position.” In: Journal of Visual Languages & Comput- ing 19.1, pp. 75–98. Nadeau D. and S. Sekine (2007). “A survey of named entity recogni- tion and classification.” In: Lingvisticae Investigationes 30.1, pp. 3– 26. Nothegger C. et al. (2009). “Selection of Salient Features for Route Directions for Route Directions.” In: Spatial Cognition and Compu- tation 5868.2004, pp. 113–136. Palacio D., C. Derungs, and R. Purves (2015). “Development and evaluation of a geographic information retrieval system using fine grained toponyms.” In: Journal of Spatial Information Science 2015.11, pp. 1–29. Papadakis E. et al. (2019). “Function-Based Search of Place Using The- oretical, Empirical and Probabilistic Patterns.” In: ISPRS Interna- tional Journal of Geo-Information 8.2, p. 92. Purves R. S. et al. (2007). “The design and implementation of SPIRIT: a spatially aware search engine for information retrieval on the In- ternet.” In: International Journal of Geographical Information Science 21.7, pp. 717–745. Pustejovsky J., J. L. Moszkowicz, and M. Verhagen (2011). “Using ISO- Space for annotating spatial information.” In: Proceedings of the International Conference on Spatial Information Theory. Randell D. A., Z. Cui, and A. G. Cohn (1992). “A spatial logic based on regions and connection.” In: Proceedings of the 3rd International Conference on Principles of Knowledge Representation and Reasoning - KR’92. Ed. by B. Nebel, C. Rich, and W. R. Swartout. Morgan Kaufmann, pp. 165–176. Raubal M. (2009). “Cognitive engineering for geographic information science.” In: Geography Compass 3.3, pp. 1087–1104. Raubal M. and W. Kuhn (2004). “Ontology-based task simulation.” In: Spatial Cognition and Computation 4.1, pp. 15–37. Raubal M. and S. Winter (2002). “Enriching Wayfinding Instructions with Local Landmarks.” In: Proceedings of the 2nd International Con- ference on Geographic Information Science. Ed. by M. J. Egenhofer and D. M. Mark. Springer, pp. 243–259. Recchia G. and M. M. Louwerse (2013). “A Comparison of String Similarity Measures for Toponym Matching.” In: Proceedings of The 1st ACM SIGSPATIAL International Workshop on Computational Models of Place - COMP’13. Ed. by S. Scheider et al. ACM, pp. 54– 61. Relph E (1976). “Place and placelessness.” In: London: Pion.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 176 bibliography
Renz J. (2002). Qualitative Spatial Reasoning with Topological Information. Vol. 2293. Springer-Verlag. Retz-Schmidt G. (1988). “Various views on spatial prepositions.” In: AI Magazine 9.2, p. 95. Richter D. et al. (2012). “How people describe their place: Identifying predominant types of place descriptions.” In: Proceedings of the 1st ACM SIGSPATIAL International Workshop on Crowdsourced and Volunteered Geographic Information - GEOCROWD’12, pp. 30–37. Richter D. et al. (2013). “Granularity of locations referred to by place descriptions.” In: Computers, Environment and Urban Systems 41, pp. 88–99. Richter K.-F. and S. Winter (2014). Landmarks: GIScience for Intelligent Services. Springer Science & Business. Ripley B. D. (1976). “The Second-Order Analysis of Stationary Point Processes.” In: Journal of Applied Probability 13.2, p. 255. Roberts K., C. A. Bejan, and S. Harabagiu (2010). “Toponym dis- ambiguation using events.” In: Proceedings of the 23rd Interna- tional Florida Artificial Intelligence Research Society Conference. Ed. by H. W. Guesgen and R. C. Murray. Vol. 10, pp. 271–276. Roller S. et al. (2012). “Supervised Text-based Geolocation Using Lan- guage Models on an Adaptive Grid.” In: Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Process- ing and Computational Natural Language Learning. Ed. by J. Tsujii, J. Henderson, and M. Pasca. Association for Computational Lin- guistics, pp. 1500–1510. Scannell L. and M. Gifford (2010). “Defining place attachment: A tri- partite organizing framework.” In: Journal of Environmental Psy- chology 30.1, pp. 1–10. Schegloff E. A. (1972). “Notes on a conversational practice: Formulat- ing place.” In: Studies in social interaction 75, pp. 75–119. Scheider S. and K. Janowicz (2014). “Place reference systems.” In: Ap- plied Ontology 9.2, pp. 97–127. Scheider S. and W. Kuhn (2010). “Affordance-based categorization of road network data using a grounded theory of channel net- works.” In: International Journal of Geographical Information Science 24.8, pp. 1249–1267. Schlieder C. (1995). “Reasoning about ordering.” In: Spatial Informa- tion Theory: A Theoretical Basis for GIS - COSIT’95. Ed. by A. U. Frank and W. Kuhn. Vol. 988. Lecture Notes in Computer Science. Springer, pp. 341–349. Schockaert S. (2011). “Vague regions in Geographic Information Re- trieval.” In: SIGSPATIAL Special 3.2, pp. 24–28.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] bibliography 177
Schwering A. and M. Raubal (2005). “Spatial relations for semantic similarity measurement.” In: Perspectives in Conceptual Modeling - ER’05. Ed. by J. Akoka et al. Vol. 3770 LNCS. Springer, pp. 259– 269. Sen S. (2008). “Use of affordances in geospatial ontologies.” In: To- wards Affordance-Based Robot Control. Ed. by E. Rome, J. Hertzberg, and G. Dorffner: vol. 4760 LNAI. Springer, pp. 122–139. Shanon B. (1983). “Answers to Where-Questions.” In: Discourse Pro- cesses 6.4, pp. 319–352. Silva M. J. et al. (2006). “Adding geographic scopes to web resources.” In: Computers, Environment and Urban Systems 30.4, pp. 378–399. Skoumas G. et al. (2016). “Location Estimation Using Crowdsourced Spatial Relations.” In: ACM Transactions on Spatial Algorithms and Systems 2.2, pp. 1–23. Smith D. A. and G. Crane (2001). “Disambiguating Geographic Names in a Historical Digital Library.” In: Research and Advanced Technology for Digital Libraries - ECDL’01. Ed. by P. Constantopou- los and I. Sølvberg. Vol. 2163. Lecture Notes in Computer Science. Springer, pp. 127–136. Smith D. A. and G. S. Mann (2003). “Bootstrapping toponym classi- fiers.” In: Proceedings of the HLT-NAACL 2003 Workshop on Analy- sis of Geographic References. Vol. 1. Association for Computational Linguistics, pp. 45–49. Speriosu M. et al. (2010). “Connecting language and geography with region-topic models.” In: Proceedings of the Workshop on Computa- tional Models of Spatial Language Interpretation - COSLI’10. Spitz A., J. Geiß, and M. Gertz (2016). “So far away and yet so close: Augmenting toponym disambiguation and similarity with text- based networks.” In: Proceedings of the 3rd International ACM SIG- MOD Workshop on Managing and Mining Enriched Geo-Spatial Data - GeoRich’16. Ed. by A. Züfle, B. Adams, and D. Wu: ACM, pp. 1– 6. Strobl C. (2017). “Dimensionally Extended Nine-Intersection Model (DE-9IM).” In: Encyclopedia of GIS. Ed. by S. Shekhar, H. Xiong, and X. Zhou. Springer, pp. 470–476. Talmy L. (1983). “How language structures space.” In: Spatial Orienta- tion: Theory, Research, and Application. Ed. by H. L. PickJr. and L. P. Acredolo. Springer, pp. 225–282. Taylor P. et al. (2009). “Where’s downtown?: Behavioral methods for determining referents of vague spatial queries.” In: Geography 3.January 2010, pp. 185–204. Teitler B. E. et al. (2008). “NewsStand: A New View on News.” In: Proceedings of the 16th ACM SIGSPATIAL International Conference
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 178 bibliography
on Advances in Geographic Information Systems. Ed. by W. G. Aref, M. F. Mokbel, and M. Schneider. ACM, pp. 144–153. Tenbrink T. (2007). Space, time, and the use of language: An investigation of relationships. Vol. 36. Walter de Gruyter. Tenbrink T. (2011). “Reference frames of space and time in language.” In: Journal of Pragmatics 43.3, pp. 704–722. Tenbrink T. and W. Kuhn (2011). “A model of spatial reference frames in language.” In: Proceedings of the 10th International Conference of Spatial Information Theory - COSIT’11. Ed. by M. J. Egenhofer et al. Vol. 6899 LNCS. Springer, pp. 371–390. Tobler W. R. (1970). “A Computer Movie Simulating Urban Growth in the Detroit Region.” In: Economic Geography 46.sup1, pp. 234– 240. Tomko M and S Winter (2006). “Recursive construction of granular route directions.” In: Journal of Spatial Science 51.1, pp. 101–115. Tomko M. and S. Winter (2013). “Describing the functional spatial structure of urban environments.” In: Computers, Environment and Urban Systems 41, pp. 177–187. Tuan Y.-F. (1977). Space and Place : the perspective of experience. Univer- sity of Minnesota Press. Tversky B. (1993). “Cognitive maps, cognitive collages, and spatial mental models.” In: Spatial Information Theory: A Theoretical Basis for GIS - COSIT’93. Ed. by A. U. Frank and I. Campari. Vol. 716. Lecture Notes in Computer Science. Springer, pp. 14–24. Twigger-Ross C. L. and D. L. Uzzell (1996). “Place and identity pro- cesses.” In: Journal of Environmental Psychology 16.3, pp. 205–220. Vandeloise C. (1991). Spatial prepositions: a case study from French. Uni- versity of Chicago Press, p. 265. Vasardani M., L. Stirling, and S. Winter (2017). “The preposition at from a spatial language, cognition, and information systems per- spective.” In: Semantics and Pragmatics 10.3. Vasardani M. and S. Winter (2016). “Place properties.” In: Advancing Geographic Information Science: The Past and Next Twenty Years. Ed. by H. Onsrud and W. Kuhn, pp. 243–254. Vasardani M., S. Winter, and K. F. Richter (2013). “Locating place names from place descriptions.” In: International Journal of Geo- graphical Information Science 27.12, pp. 2509–2532. Vasardani M. et al. (2013). “From descriptions to depictions: A concep- tual framework.” In: Proceedings of the 11th International Conference of Spatial Information Theory - COSIT’13. Ed. by T. Tenbrink et al. Vol. 8116 LNCS. Lecture Notes in Computer Science. Springer, pp. 299–319.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] bibliography 179
Wallgrün J. O., A. Klippel, and M. Karimzadeh (2015). “Towards Con- textualized Models of Spatial Relations.” In: Proceedings of the 9th Workshop on Geographic Information Retrieval - GIR ’15. Ed. by R. S. Purves and C. B. Jones. ACM, pp. 3–4. Wallgrün J. et al. (2008). “Qualitative Spatial Representation and Rea- soning in the SparQ-Toolbox.” In: Spatial Cognition V: Reasoning, Action, Interaction. Ed. by T. Barkowsky et al. Vol. 4387. Springer, pp. 39–58. Warren W. H. et al. (2017). “Wormholes in virtual space: From cogni- tive maps to cognitive graphs.” In: Cognition 166, pp. 152–163. Wasserman S. and K. Faust (1994). Social network analysis: Methods and applications. Cambridge university press. Weerdenburg D. van et al. (2019). “Where to go and what to do: Extracting leisure activity potentials from Web data on urban space.” In: Computers, Environment and Urban Systems 73, pp. 143– 156. West D. B. and Others (2001). Introduction to graph theory. Prentice hall Upper Saddle River. Wiebrock S et al. (2000). “Inference and visualization of spatial rela- tions.” In: Spatial Cognition II: Integrating Abstract Theories, Empir- ical Studies, Formal Methods, and Practical Applications. Ed. by C. Freksa et al. Vol. 1849. Springer, pp. 212–224. Wieczorek J., Q. Guo, and R. J. Hijmans (2004). “The point-radius method for georeferencing locality descriptions and calculating associated uncertainty.” In: International Journal of Geographical In- formation Science 18.8, pp. 745–767. Wing B. P. and J. Baldridge (2011). “Simple supervised document ge- olocation with geodesic grids.” In: Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Lan- guage Technologies. Ed. by D. Lin, Y. Matsumoto, and R. Mihalcea. Association for Computational Linguistics, pp. 955–964. Wing B. and J. Baldridge (2014). “Hierarchical Discriminative Classi- fication for Text-Based Geolocation.” In: Proceedings of the Con- ference on Empirical Methods in Natural Language Processing - EMNLP’14. Ed. by A. Moschitti, B. Pang, and W. Daelemans, pp. 336–348. Winter S. and C. Freksa (2012). “Approaching the notion of place by contrast.” In: Journal of Spatial Information Science 5.5, pp. 31–50. Winter S., W. Kuhn, and A. Krüger (2009). “Guest editorial: Does place have a place in geographic information science?” In: Spatial Cognition and Computation 9.3, pp. 171–173. Winter S. and M. Truelove (2013). “Talking about place where it mat- ters.” In: Cognitive and Linguistic Aspects of Geographic Space. Ed.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] 180 bibliography
by M. Raubal, D. M. Mark, and A. U. Frank. 199629. Springer, pp. 121–139. Winter S. and Y. Wu (2009). “Intelligent spatial communication.” In: Research Trends in Geographic Information Science. Ed. by G. Navratil. 199069. Springer, pp. 235–250. Winter S. et al. (2010). “Spatially Enabling ’Place’ Information.” In: Spatially Enabling Society: Research, Emerging Trends and Critical As- sessment. Ed. by A. Rajabifard. GSDI Association, pp. 55–68. Winter S. et al. (2016). “Place knowledge as a trans-disciplinary re- search challenge for Geographic Information Science.” In: Proceed- ings of the UCGIS Symposium. Ed. by J. Mennis. Winter S. et al. (2018). “A Graph Representation for Verbal Indoor Route Descriptions.” In: Proceedings of the 11th International Con- ference on Spatial Cognition. Springer, pp. 77–91. Wolter D and J. H. Lee (2010). “Qualitative reasoning with directional relations.” In: Artificial Intelligence 174.18, pp. 1498–1507. Wolter D. (2009). “SparQ-A Spatial Reasoning Toolbox.” In: AAAI Spring Symposium: Benchmarking of Qualitative Spatial and Tempo- ral Reasoning Systems, p. 53. Wolter D. and J. Wallgrün (2012). “Qualitative spatial reasoning for applications: New challenges and the SparQ toolbox.” In: Qualita- tive Spatio-Temporal Representation and Reasoning: Trends and Future Directions. Ed. by S. M. Hazarika. IGI Global, pp. 1–31. Wolter D. and M. Yousaf (2018). “Context and vagueness in au- tomated interpretation of place description: A computational model.” In: Proceedings of Workshops and Posters at the 13th Inter- national Conference on Spatial Information Theory - COSIT’17. Ed. by P. Fogliaroni, A. Ballatore, and E. Clementini: 198809. Springer, pp. 137–142. Woodru A. G. and C. Plaunt (1994). “GIPSY: Georeferenced Infor- mation Processing SYstem.” In: Journal of the American Society for Information Science 45.9, pp. 645–655. Yao X. and J. C. Thill (2005). “How far is too far? - A statistical ap- proach to context-contingent proximity modeling.” In: Transac- tions in GIS 9.2, pp. 157–178. Yao X. and B. Van Durme (2014). “Information Extraction over Struc- tured Data: Question Answering with Freebase.” In: Proceedings of the 52nd Annual Meeting of the Association for Computational Lin- guistics - ACL’14. Vol. 1, pp. 956–966. Zhang X. et al. (2012). “Disambiguating road names in text route de- scriptions using Exact-All-Hop Shortest Path algorithm.” In: Fron- tiers in Artificial Intelligence and Applications. Ed. by L. D. Raedt et al. Vol. 242, pp. 876–881.
[ April 29, 2019 at 20:02 – classicthesis version 5 ] bibliography 181
Zimmermann K. and C. Freksa (1996). “Qualitative spatial reasoning using orientation, distance, and path knowledge.” In: Applied In- telligence 6.1, pp. 49–58. Zwarts J. and Y. Winter (2000). “Vector space semantics: A model- theoretic analysis of locative prepositions.” In: Journal of Logic, Language and Information 9.2, pp. 169–211.
[ April 29, 2019 at 20:02 – classicthesis version 5 ]
Minerva Access is the Institutional Repository of The University of Melbourne
Author/s: Chen, Hao
Title: A place graph database as a qualitative human place knowledge base
Date: 2019
Persistent Link: http://hdl.handle.net/11343/222420
File Description: Thesis
Terms and Conditions: Terms and Conditions: Copyright in works deposited in Minerva Access is retained by the copyright owner. The work may not be altered without permission from the copyright owner. Readers may only download, print and save electronic copies of whole works for their own personal non-commercial use. Any use that exceeds these limits requires permission from the copyright owner. Attribution is essential when quoting or paraphrasing from these works.