M´ario Amado Alves
Adaptive Hypertext. The shattered document approach
Departamento de Ciˆenciade Computadores da Faculdade de Ciˆenciasda Universidade do Porto 2013-03-13 ii M´ario Amado Alves
Adaptive Hypertext. The shattered document approach
Tese submetida `aFaculdade de Ciˆenciasda Universidade do Porto para obten¸c˜aodo grau de Doutor em Ciˆenciade Computadores
Departamento de Ciˆenciade Computadores da Faculdade de Ciˆenciasda Universidade do Porto
2013-03-13 iv To the memory of my parents Irene and Marius. vi Abstract
We study how adaptive hypertext may improve the utilization of large online docu- ments. We put forth the inter-related concepts of shattered documents, and renoding: splitting a document into components smaller than the page, called noogramicles, and creating each page as a new assemblage of noogramicles each time it is accessed. The adaptation comes from learning the navigation patterns of the usors (authors and readers), and is manifested in the assemblage of pages. Another essential trait of our work is the utilization of user simulation for testing our hypotheses. We have created software simulators and conducted experiments with them to compare several adaptive and non-adaptive configurations. Yet another important aspect of our work was the study and adoption of the technique of spreading activation to explore the network database of the learnt model of travels. We have realised a quantitative evaluation based on utilization quality measures adapted to the problem: session size, session cost.
vii viii Resumo
Estudamos como o hipertexto adaptativo pode melhorar a utiliza¸c˜aode documen- tos em-linha de grande dimens˜ao. Apresentamos os conceitos interrelacionados de document fragmentado e rela¸cagem (recria¸c˜aode n´os):separa¸c˜aodo documento em componentes de dimens˜aoinferior `ap´agina,chamados noogram´ıculos, e cria¸c˜aode cada p´aginacomo uma nova montagem de noogram´ıculosde cada vez que ´eacedida, deste modo criando novos n´osna rede hipertextual, `amedida que a utiliza¸c˜aoprogride. A adapta¸c˜aoprov´emde aprender os padr˜oesde navega¸c˜aodos utilizadores (autores e leitores) e manifesta-se na montagem das p´aginas. Outro tra¸coessencial do nosso trabalho ´eo recurso `a simula¸c˜ao para testar as nossas hip´oteses.Cri´amossimuladores em software e realiz´amosexperiˆenciascom eles para comparar v´ariasconfigura¸c˜oes adaptativas e n˜ao-adaptativas. Ainda outro aspeto importante deste trabalho ´eo estudo e ado¸c˜aoda t´ecnicade propaga¸c˜aoda ativa¸c˜ao para explorar a base-de-dados reticular do modelo de viagens aprendidas. Realiz´amosuma avalia¸c˜ao quantitativa baseada em medidas de qualidade da utiliza¸c˜aoadaptadas ao problema: tamanho da sess˜ao,custo da sess˜ao.
ix x R´esum´e
On ´etudiecomme l’hypertexte adaptatif peut am´eliorerl’utilisation de documents en- ligne de grande dimension. On pr´esente les concepts, en rapport entre eux-m`emes,du document fragment´e et le rela¸cage (recr´eationde noeuds): la s´eparationdu document en des composants de dimension inf´erieure `acelle de la page, nom´esles noogramicules, et la cr´eationde chaque page comme un nouvel assemblage de noogramicules `achaque fois qu’on l’acc`ede, cr´eant ainsi de nouveaux noeuds sur le r´eseauhypertextuel, au fur et `amesure du progr`esde l’utilisation. L’adaptation vient de l’apprentissage des mod`eles de navigation des utilisateurs (auteurs et lecteurs) et elle se r´ev`ele`atravers l’assemblage des pages. Un autre trais essentiel de notre travail, c’est le recours `a la simulation afin de tester nos hypoth`eses. On cr´eades simulateurs logiciels et on r´ealisades exp´eriencesavec eux pour comparer plusieurs configurations adaptatives et non adaptatives. Un autre aspect important de ce travail c’est l’´etudeet l’adoption de la technique de propagation de l’activation pour exploiter la base de donn´esr´eseau du mod`elede voyages appris. On r´ealizaune ´evaluation quantitative support´eepar des mesures de qualit´ed’utilisation, adapt´eesau probl`eme:taille de s´ession,coˆutde s´ession.
xi xii Acknowledgements
I love deadlines. I like the whooshing sound they make as they fly by. Douglas Adams
I am extremely indebted to my advisor Doctor Al´ıpioJorge and co-advisor Doctor Z´e Paulo Leal.
Al´ıpiois the most wise person I know. Each of his numerous advice was always entirely pertinent and convenient. Had I followed them all and this thesis would have been a work of absolute perfection. Al´ıpio’ssupport of my work stood unabated through the unending stream of deadline missing after deadline missing from my part. I was as surprised as I was thankful for this continued support. I was surprised because Al´ıpio is the most wise person I know, and I would expect even a mildly clever person to spot a clear lost case and hastily detach themselves from the dead weight. Paradoxically, the fact that you are now reading these lines on an acceptable if imperfect thesis is a result of Al´ıpiobeing the most wise person I know.
I am indebted to Professor Pavel Brazdil for being the excellent leader of LIAAD - INESC Porto (formerly NIAAD - LIAAC), the laboratory where this whole business started.
I am indebted to Rodolfo Matos, the prolific sysadmin of LIAAD, for his prompt support and enduring friendship.
I am grateful to all NIAAD and DCC members for their support and comradeship.
I am indebted to the Funda¸c˜aopara a Ciˆenciae Tecnologia for supporting four years of doctoral research.
I am indebted to Ada Europe, APPIA, Prolearn, and Universidade Aberta, for conference-
xiii going support.
I am grateful to Cec´ılia.
With the grace of God.
xiv Contents
Abstract vii
Resumo ix
R´esum´e xi
Acknowledgements xiii
List of Tables xxiii
List of Figures xxvii
1 Introduction 1
1.1 Context ...... 1
1.1.1 Rationale for hypertextualization ...... 2
1.1.2 Limitations of hypertext ...... 3
1.2 Solutions ...... 4
1.2.1 Our solution ...... 6
1.2.2 Hypotheses of this work ...... 6
1.3 Main contributions ...... 7
1.4 Structure of the text ...... 8
1.4.1 Notation ...... 9
xv 2 Hypertext 11
2.1 Definition ...... 11
2.2 Terminology ...... 12
2.3 Forms of hypertext ...... 13
2.3.1 The words used ...... 19
2.3.2 Items vs. connections ...... 20
2.3.3 The Monographic Principle ...... 22
2.3.4 Summary of hypertext history ...... 24
2.3.5 The document dogma ...... 24
2.3.6 Information search—the impossible that is done ...... 25
2.3.7 Aporias of adaptation ...... 26
2.3.8 Minor issues ...... 27
2.4 Structure of documents ...... 36
2.4.1 Traditional document structure ...... 36
2.4.2 Standard hypertextualization ...... 37
2.5 Learning systems ...... 40
2.5.1 Adaptive hypertext techniques ...... 41
2.5.2 Learning Systems highlights ...... 44
2.6 Summary ...... 46
3 A new model for adaptive hypertext 49
3.1 Motivation ...... 49
3.1.1 Information, not documents ...... 49
3.1.2 Guidelines for adaptive hypertext ...... 50
3.2 Model design ...... 54
3.2.1 The Shattered Documents model ...... 54
xvi 3.2.2 Adaptive information, and author as first reader ...... 56
3.2.3 Interface design ...... 57
3.2.4 Detailed design with a network data model ...... 60
3.3 Techniques and tools reused ...... 60
3.3.1 A unified model of spreading activation ...... 63
3.3.2 A didactical example ...... 63
3.3.3 Benefits of spreading activation for information retrieval . . . . 66
3.3.4 The generic model ...... 67
3.3.5 About the implementation ...... 70
3.3.6 Leaky Capacitor Model (LCM) ...... 71
3.3.7 Reverberative Circles (RC) ...... 73
3.3.8 Waterline ...... 74
3.4 Algorithms ...... 74
3.4.1 Overview ...... 75
3.4.2 Formalization ...... 76
3.4.3 Start page algorithms ...... 76
3.4.4 Recentring algorithms ...... 77
3.4.5 Learning algorithms ...... 79
3.5 Summary ...... 80
4 Experimental methodology 81
4.1 Simulation ...... 82
4.1.1 Formalization ...... 83
4.2 Experiments ...... 84
4.3 Parameter settings ...... 85
4.3.1 The document ...... 88
xvii 4.3.2 One thousand nodes ...... 88
4.4 Evaluation methodology and measures ...... 91
4.4.1 Session size ...... 91
4.4.2 Session cost ...... 91
4.5 Statistics ...... 92
4.5.1 Common and bottom line statistics ...... 94
4.5.2 Statistics in the Outcomes tables ...... 95
4.5.3 Statistics in the Results tables ...... 96
4.5.4 Statistics in the Evolution tables ...... 96
4.5.5 Alternate terms ...... 97
4.6 Summary ...... 97
5 Results 99
5.1 Main results ...... 100
5.1.1 Session size and success rate ...... 101
5.1.2 Evolution ...... 102
5.2 Testing link types ...... 103
5.3 Testing adaptative techniques ...... 105
5.4 Summary ...... 105
6 Conclusions 107
6.1 Main conclusion ...... 107
6.2 The pros and cons of simulating ...... 107
6.3 Paths not taken ...... 108
6.4 Future work ...... 109
A Electronic archive 127
xviii B Program listings 129
B.1 Package Arm05 Model (spec) ...... 129
B.2 Package Arm05 Model (body) ...... 131
B.3 Procedure Arm05 Model.Get Info (body only) ...... 134
B.4 Package Kasim2 (spec) ...... 137
B.5 Package Kasim2 (body) ...... 143
B.6 Package Kasim2.Activation (spec) ...... 151
B.7 Package Kasim2.Activation (body) ...... 152
B.8 Package Kasim2.Comparate (spec) ...... 160
B.9 Package Kasim2.Comparate (body) ...... 161
B.10 Procedure Kasim2.Comparate.Experiment (body only) ...... 186
B.11 Package Kasim2.Markov (spec) ...... 187
B.12 Package Kasim2.Markov (body) ...... 187
B.13 Package Kasim2.Markov With Heuristics (body) ...... 190
B.14 Package Kasim2.Markov With Heuristics (body) ...... 193
B.15 Package Kasim2.Structural (spec) ...... 197
B.16 Package Kasim2.Structural (body) ...... 197
C Supplemental Items 201
C.1 Software requirements for the Knowledge Atoms design ...... 201
C.2 Kasim 1. First cycle of experiments ...... 202
C.2.1 Second round: 60 atoms, 10 oracles ...... 205
C.2.2 Conclusion of cycle one ...... 207
C.3 Notes on Network Data Models ...... 207
C.3.1 Primary concepts of network data structures ...... 208
C.3.2 The untyped network hypothesis ...... 209
xix C.3.3 Other designs ...... 214
C.3.4 A network calculus ...... 215
C.3.5 Fundamental entities, definitions ...... 215
C.3.6 Useful theorems ...... 216
C.3.7 Proofs, lemmas, axioms ...... 217
D Detailed results of experiments 219
D.0.8 Common and bottom line statistics ...... 221
D.0.9 Statistics in the Outcomes tables ...... 222
D.0.10 Statistics in the Results tables ...... 223
D.0.11 Statistics in the Evolution tables ...... 223
D.0.12 Alternate terms ...... 224
D.1 Complete results of configuration Shattered Document ...... 225
D.2 Complete results of configuration Shattered Document with Random Clicks ...... 233
D.3 Complete results of configuration Shattered Document with Random Pages...... 241
D.4 Complete results of configuration Shattered Document with Random Learning ...... 249
D.5 Complete results of configuration Shattered Document with Random Pages and Random Learning ...... 257
D.6 Complete results of configuration Shattered Document with Random Document (fixed) ...... 265
D.7 Complete results of configuration Structural ...... 273
D.8 Complete results of configuration Shattered Document with high max- imum cost ...... 281
D.9 Complete results of configuration Structural with high maximum cost . 289
xx D.10 Complete results of configuration Shattered Document with Child links only ...... 297
D.11 Complete results of configuration Shattered Document, no Next links, weak Residual ...... 305
D.12 Complete results of configuration Shattered Document, weak Next links 313
D.13 Complete results of configuration Shattered Document with Strong Legacy ...... 321
D.14 Complete results of configuration Shattered Document, dense start . . 329
D.15 Complete results of configuration Shattered Document with Markov Chains ...... 337
D.16 Complete results of configuration Shattered Document with Markov Chains with heuristics ...... 345
D.17 Complete results of configuration Shattered Document, Trimmed Tree . 353
D.18 Complete results of configuration Structural, Trimmed Tree ...... 361
D.19 Complete results of configuration Shattered Document with Markov Chains, Trimmed Tree ...... 369
D.20 Complete results of configuration Shattered Document with Markov Chains with heuristics, Trimmed Tree ...... 377
xxi xxii List of Tables
2.1 Summary of hypertext history ...... 24
4.1 Some scalar parameters of the simulator ...... 87
4.2 Distribution of nodes per level in the ARM ...... 88
4.3 Random sample of sections ...... 90
4.4 Distribution of nodes per level in the Cut ARM ...... 90
4.5 Distribution of nodes per level in the Trimmed ARM ...... 91
5.1 Main results...... 100
5.2 P-values for the main results...... 101
5.3 Results for maximum cost = 6...... 102
5.4 Varying the link weights ...... 104
5.5 Markov chain results ...... 105
xxiii xxiv List of Figures
1.1 Adaptive hypertext process...... 2
1.2 Example of online shop recommendations...... 5
2.1 Screenshot of a Xanadu prototype on Nelson 2007a...... 15
2.2 Illustration of an advanced form of hypertext on Nelson 2007b...... 16
2.3 SWI-Prolog help panel...... 17
2.4 SWI-Prolog help panel for consult/1 ...... 18
2.5 Dictionary...... 19
2.6 SWI-Prolog help for catch/3...... 28
2.7 Alternative to the design in figure 2.6 ...... 28
2.8 The web interface of the network router Linksys WRT54G...... 29
2.9 How compound terms are (mis)treated in Dictionary...... 30
2.10 The Back button aporia. In certain contexts the user expects a chrono- logical behaviour. 33
2.11 The Back button aporia. Why the Back button should not behave chronologically. 33
2.12 Standard hypertextualization of the sequential structure...... 39
2.13 Real look of the first page in figure 2.12...... 39
2.14 The Knowledge Sea interface (KnowledgeSea)...... 46
xxv 2.15 Detail of The Knowledge Sea interface (KnowledgeSea)...... 47
3.1 Model of the same document in figure 2.12 but with the shattered document approach and the two types of connection Next (N) and Child (C)...... 55
3.2 Page made up of document fragments...... 55
3.3 Example design of our own adaptation model...... 59
3.4 Relating the web page as seen and the graph model underneath. . . . . 61
3.5 Travelling ...... 62
3.6 Toy network, for a general understanding of spreading activation . . . . 64
3.7 Instances of spreading activation over the toy network (transcripts of sessions with the Minibrain program)...... 65
3.8 Step algorithm...... 70
3.9 Minibrain usage...... 70
3.10 Super Page algorithm ...... 76
3.11 Main recentring algorithm ...... 77
3.12 Pure Markov chains recentring algorithm ...... 78
3.13 Heuristical Markov chains recentring algorithm ...... 78
4.1 Choose algorithm ...... 84
4.2 Evolution of session size in an exploratory experiment ...... 85
4.3 Top level models and respective components of the experimental setup. 86
4.4 Partial configuration map of the simulator...... 87
4.5 Hierarchical structure of the ARM (excluding annexes)...... 89
5.1 Session cost evolution for the shattered document D.9 ...... 103
5.2 Session costs and sizes for the original structure D.8 ...... 103
xxvi C.1 Hierarchical structure of document X3 (CHILD links) ...... 203
C.2 Complete structure of document X3 ...... 204
C.3 Evolution of session size for each oracle ...... 205
C.4 Hierarchical structure of document X4 (CHILD links) ...... 206
C.5 Evolution of session size ...... 207
C.6 A subset of XQuery use case 1.1.2 (Chamberlin et al. 2005) represented in (a) the original XML format, (b) a typed network a la RDF. This example illustrates various issues discussed in the text...... 209
C.7 The direct attribute structure equates a connection...... 211
C.8 Untyped base structure to represent an explicit, extensively defined set
X = {x1, . . . , xm}...... 212
C.9 The road intersection pattern (a), for representing an (abstract) typed connection (b), in the untyped network base...... 212
C.10 A subset of XQuery use case 1.1.2 (Chamberlin et al. 2005) represented in: (a) the original XML format; (b) a typed network; (c) the untyped network base. This example illustrates various issues discussed in the text...... 213
xxvii xxviii Chapter 1
Introduction
Traveller, there are no paths. Paths are made by walking. Ant´onioMachado
1.1 Context
This thesis studies the process of hypertextualization and hypertext adaptation, as a means to improve the utilization of large online documents. Figure 1.1 illustrates this two-phase transformation process from a traditional document, or set of items, to hypertext, and therefrom to a different, adapted, hypertext—hopefully a better one. The diagram depicts the overall process; the techniques named therein are just a few of many possible ones, all addressed later.
We study the different parts of the process. In general, we delve more deeply into the adaptive parts of the process, situated on the right-hand half of the diagram. Our main original contributions are the concept of renoding, or shattered documents, and a simulator of user behaviour, which we use to automate the evaluation of different hypertext systems.
In sum, we study the large online document, and how its utilization might be improved by means of adaptive hypertext features. Large means an extent such that the document cannot be seen all at once. In other words: large = (much) larger than a screenful.
1 2 CHAPTER 1. INTRODUCTION
Figure 1.1: Adaptive hypertext process. The horizontal direction represents transformation. Vertical arrows represent data flow. See text.
1.1.1 Rationale for hypertextualization
It is a fair assumption that large online documents have much to gain from being structured as hypertext. The amount of hypertext present in the Word “Wild” Web, its sheer weight, supports this assumption. (However its relative weight is small, 44% only, against 56% of PDF documents. More on this later.)
The collective mind of Web authors, manifested in the authored Web links, has also been shown a useful resource for exploring the Web space. The current epitome of this approach is the famous PageRank algorithm, which arranges Google results (Page et al. 1999).
Since the Web can be seen as a huge hypertext document, we may transpose these facts to the general concept of large document, and hypothesise that any large online document will gain by taking the form of hypertext and being adapted to—and by—its users and authors, or usors1.
More generally, hypertextualization can be seen as an effect of the desire to augment intelectual power via technology (Bush 1945). This movement is pervasive to all mankind, and clearly justifiable. This rationale of intelect augmentation is adressed lengthly in chapter 2.
1After having arrived at the concept and term of usor, we have found the term wreader being used for a similar concept, as reported by Weel 2006. We have stayed with usor: its morphology is more regular than that of wreader, and its sound is more distinguishable from user than the sound of wreader from reader. 1.1. CONTEXT 3
1.1.2 Limitations of hypertext
The main problem with hypertext is an increased utilization effort relative to normal text. This accrued effort is two-fold: there is an increased effort of authoring, and there is an increased effort of navigation.
1.1.2.1 Hypertext is costly to build
There is an increased authoring effort of creating the pages and the links—accrued relatively to the work of producing normal text (Bra et al. 1999). The hypertext author has to organize or divide the content into pages, and create links between the pages. The number of possible page partitions and manner of linking is virtually infinite, which puts upon the author the burden of making an innordinate number of decisions that are difficult, violent (Belo 1991)—and therefore very costly.
One might think that normal text has a specific cost of its own too: because of the sequential nature of normal text, special devices of anaphor, narrative, prosody, punctuation, textual cohesion—rhetorical devices in whole—must come into play to make for a good sequential reading. But it has been shown that such factors play their part in hypertext too (Mancini 2004). For one, hypertext is still text. Even if the pages are short, they are still large enough for most rhetorical devices to operate in—and be required to.
1.1.2.2 Hypertext is costly to navigate
There is an increased reading effort of navigating an hypertext—additional to the effort of reading normal text. Essentially, the user is put in charge of constructing their own unique pathway through a variety of options (Lawless et al. 2003, Raskin 2000). This activity incurs the extra cognitive cost of making choices, from a sometimes unwiedy number of options, of links on a page.
A related, often purported limitation of hypertext is the lost in hyperspace syndrome. We discuss this on chapter 2, particularly on section 2.3.8.6. 4 CHAPTER 1. INTRODUCTION 1.2 Solutions
A number of automatic adaptive techniques have been tried or proposed as a means to alleviate the costs of both authoring and reading hypertext (Bra et al. 1999). Such techniques are analysed at length on chapter 2, particularly on section 2.5. Most if not all such proposals revolve around the concept of relinking—creating or changing links automatically, with the objective of helping users in their navigation. Those proposals are academic, stemming mostly from the Intelligent Tutoring Systems area, and are still confined to the laboratory. So is our own proposal.
In the real world, there is Google and Amazon. These resources can be seen as large hypertexts too, because they have pages and links. Google results are automatically created links from a search expression to related items on the Web. Amazon recom- mendations are automatically created links to related items on the store. Amazon and similar sites are adaptive hypertexts.
The principal part of Google algorithm is the PageRank algorithm that order the results (Page et al. 1999). PageRank relies entirely on the authoring effort of others, namely of all authors of all the pages on the Web, because it draws results from the analysis of the links between the pages, and such links and pages were created at a cost to the authors, as we have discussed. So PageRank is not really a solution to the hypertextualization cost problem, it just moves the problem out of its sight. Without authored links, there would be no Google.
The recommendations at Amazon, or at other similar online shops, are computed based on the recorded navigation data, using a technique called collaborative filtering (Goldberg et al. 1992). The navigation data includes conversion information: wheather the user has bought the item or not. See figure 1.2 for an example; note the recomenda- tions at the bottom of the large central zone; note the recommendations clearly marked as derived from the choices made by other users: “Customers who were interested in Startone CG 851 4/4 bought the following products”, and the items recommended by some other means, simply titled “related products”.
The recorded navigation data represent the actions made by users. Using this data is indeed likely to provide a means of alleviating the hypertextualization effort. Actually our own proposal in this thesis can be seen as a variant of this approach, as we shall see. That is, our system can partially be seen as a recommendation system. Among other similarities of our proposal with online shop recommendations, there is the obvious alignment between the concepts of conversion (online shop) and successful session 1.2. SOLUTIONS 5
Figure 1.2: Example of online shop recommendations. 6 CHAPTER 1. INTRODUCTION
(our system).
Recommendation systems have yielded good results in online shopping—we have not seen the method applied to documents. The laboratorial, academic adaptive tech- niques (cf. chapter 2, particularly section 2.5), have indeed been aplied to documents, with results that are also, so far, generally positive. However, it is hard to draw convincing comparisons because there is no common methodology, no common mea- sures of hypertext usability used across the different proposals. We wanted to improve on this situation. In particular, we wanted to explore the following observation: all existing systems, laboratorial or real, treat the pages as given, integral wholes.
1.2.1 Our solution
We questioned the presumed atomicity of pages, on the assumption that it might be an obstacle to better adaptation. Our rationale is based greatly on the observation that when we consult a document for reference, we end up selecting a small part of the document to satisfy our precise information need. So perhaps we could split, or shatter, a document into such small parts, or noogramicles, in order to adapt pages made of them towards the (computed) needs of the reader.
So, whereas in all existing systems (laboratorial or real) the pages are given, integral wholes, in our system the pages, or nodes, are an assemblage of small document parts, which we call noogramicles. Some noogramicles are shown contracted and actually constitute links to their fully expanded expression, or view. So a page in our system is actually a set of items of two types: document parts, links to other document parts.
Each page is created anew upon each request. So we have called this approach renoding, as a paraphrase of relinking. The constituend noogramicles of each page are selected based on usage, or navigation data, as in recommendation systems. This selection is done using a spreading activation algorithm, described later.
1.2.2 Hypotheses of this work
Naturally we expect our approach to improve upon a non-adaptive document, and so our main hypothesis is expressed thus:
Shattered Document Hypothesis (main hypothesis). The navigation effort of 1.3. MAIN CONTRIBUTIONS 7
a user in a shattered adaptive hypertext document is reduced with respect to the original hypertext document.
We also hope our spreading activation algorithm to fare better than a standard adaptive technique like Markov chains.
Spreading Activation Hypothesis (supplemental hypothesis). Spreading acti- vation is a better adaptation model for shattered documents than first order Markov chains.
Finally, a part of our method consists in treating all relations, or links, equally. Namely, we propose that, for adaptation purposes, the original document relations between noogramicles, namely the relations of sequence, hierarchy, and cross-reference, equate each other and, more importantly, they equate the relation established by the user when travelling between noogramicles. This amounts to viewing the author simply as the first user, or usor of the document, i.e. we unify the concepts of author and reader.
Usor Hypothesis (supplemental hypothesis). Varying the weights of links ac- cording to their type does not have an influence on results.
1.3 Main contributions
The main contributions of this thesis are concentrated in the new concept of renoding. Renoding is a paraphrase of relinking. Relinking is the adaptive hypertext state-of- the-art item that consists of changing, creating, or deleting hyperlinks (Bollen 2001). A foundational observation of our research is that there is no a priori reason why links should be thus changed and pages not. In particular, we observe that the state-of-the- art tacitly subsumes pages as given, integral items.
Renoding is necessarily achieved by means of shattering the document into small pieces, or noogramicles, for adaptation. It is very important to keep in mind that renoding and shattering are done for the purpose of adaptation only. That is, renoding is strictly an adaptation technique.
Given the extremely networked nature of the data structures involved, we have decided to use spreading activation as a technique to explore these data graphs. The use of 8 CHAPTER 1. INTRODUCTION spreading activation for data analysis is still an evolving research issue, therefore our findings in this area are also a contribution of this thesis.
As an experimental test bed, we have built a user simulator. Although potentially imperfect and necessarily simplified with respect to reality, the simulator enables extensive testing with different variants and can be regarded as a contribution per se. It also provides a well defined characterization of the assumptions under which the proposed approaches can operate. The evaluation of selected techniques, as well as the callibration of the simulator, on real user experiments has been left for future work. Using the proposed simulator, we have compared our approach against non- adaptive variants, and also against adaptive variants using a standard recommendation algorithm based on Markov Chains. Our approach has indeed fared the best in our main measure of session cost, ultimately verifying all our hypotheses.
1.4 Structure of the text
This thesis is organized into the following main parts.
Chapter 1. Introduction. Main theme and hypotheses briefly presented and mo- tivated.
Chapter 2. Hypertext. Discussion of hypertext in general, its history, including background and related work, and further motivation for the approaches in this thesis.
Chapter 3. A new model for adaptive hypertext. Description of the shattered document approach, including algorithms.
Chapter 4. Experimental methodology. Definition of the experimental setup, in- cluding simulation algorithms.
Chapter 5. Results Presentation of the results of the experiments.
Chapter 6. Conclusions Conclusions and future work.
References. Bibliographical descriptions of cited or studied works, websites, software works.
Appendix A. Electronic archive. DVD-ROM or card with the software created in this thesis, or indication of where on the web to obtain it. 1.4. STRUCTURE OF THE TEXT 9
Appendix B. Program listings. Listings of the principal source code units cre- ated.
Appendix C. Supplemental writings. A few writings that did not fit on the main body of the thesis, including a description of the first, small scale, exploratory experiments and simulator.
Appendix D. Detailed results of experiments. Complete result data of the large scale experiments reported, in the form of tables and charts.
Physical copies of this thesis are normally divided in volumes (the page numbers and quantities are approximate):
Volume Page numbers Qt. of pages I. Main text and appendix A i–xxv, 1–125 150 II. Appendices B–D 125–375 250 total = 400
1.4.1 Notation
Forms like Rome 2003 are bibliographical references. Forms like Turbo C, undated, are references to entities of a different type than documents, but still referable, like software works or websites. All references are collected and described on the respective section.
Single word, or expression, quotations take this form. Short, in-text quotations “take this form”. Single word or expression quotations take the same form as emphasis, not to overload the text with too many distinct notations. All short or long quotations are accompanied by the respective bibliographical reference.
“A long quotation takes the form of this paragraph. A long quotation takes the form of this paragraph. A long quotation takes the form of this paragraph.” 10 CHAPTER 1. INTRODUCTION Chapter 2
Hypertext
Our writing equipment takes part in the forming of our thoughts. F. Nietzsche
In this chapter we define hypertext, and discuss the aporias, or theoretical difficul- ties, of text, hypertext, and adaptive hypertext. Occasionally, we lay down certain assumptions, which we have found necessary in order to proceed with our work. We also chart the history of hypertext in this chapter.
2.1 Definition
An entirely consensual definition of hypertext does not exist. Below (section 2.3) we shall discuss this issue in detail and arrive at a working definition, which we preview here for convenience:
Hypertext is an interface to interconnected items of textual or pictorial information, which interface lets the user follow any connection, and also records the connections followed and lets the user relive them at will, typically by means of a “back button”.
11 12 CHAPTER 2. HYPERTEXT 2.2 Terminology
We use the following non-trivial or specific terms in this thesis. Terms marked with an asterisk have a specific meaning in this thesis i.e. we give the meaning specifically used in this thesis and particulary in the shattered document design.
Atom* See noogramicle.
Codex The current form of the physical book: a set of rectangular leaves of paper bound together along one side.
Context The wider or historical circunstances. Other authors use the term user context instead of user input.
Cybernetics The science that studies the abstract principles of organization in com- plex systems. (Heylighen & Joslyn 2001)
Distal content The page at the other end of the link. (Olson & Chi 2003)
Fan-in The quantity of afferent (incoming) immediate connections from a node in a tree or graph data structure. A.k.a. in-degree.
Fanout The quantity of efferent (outgoing) immediate connections from a node in a tree or graph data structure. A.k.a. out-degree.
Information need The information need of a user of an information repository. Syn- onymous with the term and concept of reference question from library science. The information need triggers, and sustains, a search session.
Noogramicle* The smallest constituent of meaning on a document. A caption, a figure, a formula, a footnote, a heading, a listing, a paragraph, a sentence, a table, a title, etc. Synonyns used at some time in our work include atom, knowledge atom, paragraph, extended paragraph; these forms may surface for historical reasons on software items; please adjust.
Oracle* The noogramicle, or noogramicles, that constitute the answer to the refer- ence question. We chose this term over the more common target or goal, to facilitate distinction from other meanings of the latter, notably the end point of a link, or the head of a directed connection in a data graph, which are concepts operating in this thesis also. 2.3. FORMS OF HYPERTEXT 13
Page* An assemblage of noogramicles. More precisely, of views of noogramicles.
Reference question See information need.
Session A finite episode of utilization of an hypertext system, by a user with a reference question in mind. The session represents the user searching for the answer to the reference question. The session terminates upon finding the answer (successful session) or giving up (unsuccessful session).
Session cost* Cognitive effort associated with the navigation aspect of a session. A rough measure is Session size. Finer measures takes into account other factors like page size, need for scrolling, etc.
Session size Session size = number of pages visited = number of clicks + 1.
View* The view of a noogramicle is the readable representation of the noogramicle. It is either expanded or contracted. The expanded or full view represents the noogramicle entirely. The contracted view, or label, serves as the anchor of a link. For indicative noogramicles like headings and titles normally the full view is used as the label i.e. there is no contraction proper. For propositional noogramicles like paragraphs normally the first few words are used to form the label; for items with a caption like tables and figures normally the caption is used for this purpose.
2.3 Forms of hypertext
The first aporia of hypertext concerns the notion itself. Very different kinds of hypertext have been proposed, some even antagonistic between them. For example, Nelson’s conception is consistently put in dire opposition to Berners-Lee’s, or HTML. This state of affairs makes the use, the denotation, of the word hypertext inaccurate, to say the least.
In the 1960’s Ted Nelson coined the very word hypertext, with the general meaning of
“a body of written or pictorial material interconnected in such complex way that it could not conveniently be presented or represented on paper.” (Nelson 1965) 14 CHAPTER 2. HYPERTEXT
The word has thriven, as we know. In the course of these four or five decades, it has become an essential part of both popular and scientific culture. Today, ar- guably the most common species of hypertext is that inhabiting the World Wild Web. N¨urnberg et al. 1996 integrate well from several notable sources (Engelbart & English 1968, Halasz 1988, Marshall et al. 1994), for a definition of such kind of hypertext:
“Information realized in the interface by connected “pages” of text and graphics traversed through a “point-and-click” navigation mechanism.” (N¨urnberg et al. 1996)1
Surveying an even wider range of sources, Mancini 2004 arrives at a tripartite classifi- cation, of which we will attempt a synthetic formulation2
• page-based hypertext: the common, web-like hypertext
• semantic hypertext, characterized by having typed connections
• spatial hypertext: semantic hypertext visualized graphically.
Mancini 2004 stands as one of the best works we have studied on the subject of hyper- text, and one of the very few correctly acknowledging the existence of various kinds of hypertext. Nevertheless, we observe that, strangely, she colocates Ted Nelson’s hypertext alongside web hypertext, in the page-based category. This might be unfair towards Nelson’s original proposal, called Xanadu:
“I believe it was in 1968 that I presented the full 2-way Xanadu3 design to a university group, and they dismissed it as “raving”; whereupon I dumbed it down to 1-way links and only one visible window. When they asked how the user would navigate, I suggested a backtrackable stack of recently visited addresses. I believe that this dumbdown, through the various pathways of projects imitating one another, became today’s general design, and I am truly sorry for my role in it.” (Nelson 1999)
1Not surprisingly, this definition approximates closely that of dicionaries, cf. figure 2.5. 2This formulation is our own very subjective reinterpretation of Mancini 2004, because we were not able to locate, on this part of Mancini’s text, a clear statement of a sharp, essential distinction between the proposed classes. 3Xanadu is the design already presented on Nelson 1965. 2.3. FORMS OF HYPERTEXT 15
Figure 2.1: Screenshot of a Xanadu prototype on Nelson 2007a.
Xanadu presents multiple windows at the same time, and visible connections between them. See also figures 2.1 and 2.2. Clearly, spatial hypertext in Mancini’s classifica- tion, not page-based.
Incidentally, note how Nelson, according to the account above, invented the Back Button right there. Let us take notice that the associated requirement—navigational memory—has been a staple of hypertext since its inception on Bush 1945. And many other essential sources, notably Nelson 1963, indicate clearly that users are supposed to be able to backtrack to any past point in their travels.
And indeed the Back Button has proven itself a formidable weapon against the lost in hyperspace syndrom, discussed later (section 2.3.8.6. It accounts for up to 42% of user actions with web browsers (Cockburn et al. 2002). If there ever was a silver bullet of interfaces, this was it. So we must make the Back Button—or some other form of access to the navigation history of the user—a strict requirement of our working definition of hypertext. 16 CHAPTER 2. HYPERTEXT
Figure 2.2: Illustration of an advanced form of hypertext on Nelson 2007b. The name is now transliterature, but the concept is essentially the same as Xanadu and Nelson 1965. 2.3. FORMS OF HYPERTEXT 17
Figure 2.3: SWI-Prolog help panel. In the help text, the terms set in bold face and coloured green are clickable connections. For example, clicking on consult/1 (under the arrow on the picture) replaces the page with the one on figure 2.4
Above, and on other sources, the Web has been given as an example of page-based hypertext. It must be noted that page-based hypertext is also common in offline programs, for example in help systems, multimedia encyclopedias, etc. Recall pro- gramming in the early 1980’s—before the Web—with the Turbo C compiler and environment. With the cursor upon a keyword or library entity in our program text, a hit of the F1 key would bring up instant documentation about it, and these help panels contained similar linkage to others. For a current, after the Web, example of offline programming help let us refer to the help system of SWI-Prolog, cf. figures 2.3 and 2.4.
We observe that the SWI-Prolog help system, like many such systems, unfortunately, lacks the Back Button, or any other navigational memory.
For another example of offline hypertext consider the Dictionary application, depicted in figure 2.5. In Dictionary, every word is clickable, so the designers chose not to use any special typography—wisely, because that would only produce unecessary visual 18 CHAPTER 2. HYPERTEXT
Figure 2.4: SWI-Prolog help panel for consult/1
clutter. Note also that, unlike SWI-Prolog help, Dictionary features Back and Forward buttons, and is therefore an hypertext system by all accounts.
In sum, there are various kinds of hypertext. We can define the invariants as follow:
• hypertext is an interface to interconnected items
• the items are of information, textual or pictorial
• the interface lets the user follow any connection
• the interface records the connections followed, and lets the user relive them at will; in particular, the interface provides a back button
And this will be our working definition of hypertext. 2.3. FORMS OF HYPERTEXT 19
Figure 2.5: Dictionary.
2.3.1 The words used
Nelson 1965 introduces other hyper concepts: hypermedia, hyperfilm. But does not offer a rigourous ontology. The words hypermedia, hypertext are used interchangeably in the research literature i.e. they mean the same. The word hypermedia seems to be preferential in the learning systems4 area, cf. Chen & Ford 1998; it is also often used in commercial marketing, to inform potential costumers of an hypermedia product that it contains multimedia, not just plain text. The word hypertext is used everywhere else—including here.
Let me suggest that the text in hypertext, as used by Nelson and others, already means more than just plain text, more than just letters. It means document, discourse, sign. It includes pictures, graphics, formulae, even sound and films. In a word: content. This, shall we say, semiotic meaning of text is common in the humanities (an habitual Nelson dweling); recall the idiom subtext of a film, etc. Nelson’s own definition supports this approach: hypertext is ... “textual or pictorial”.5
4Comprising ITS: intelligent tutoring systems. 5Curiously enough, this extension of the meaning of text mirrors the way Otlet, as we shall see, has extended the meaning of document to include all information objects, further than books, e.g. artifacts, archeological findings, models, didactical toys, works of art, etc. Or as, in semiotics, 20 CHAPTER 2. HYPERTEXT
2.3.1.1 Literary vs. technical hypertext
Consider the two distinct classes of text, or discourse:
• literary, poetic, artistic discourse: works of fiction, poetry, etc.
• technical, scientific, utilitarian discourse:6 manuals, thesis, articles, etc.
We note that Nelson often uses the term literature for any kind of text. Here we use the terms literary and technical for the two classes purportedly in opposition.
This aporia relates closely to rhetorical categories. Literary hypertext studies seem to proceed in terms of rhetorical categories, cf. Mancini 2004.
Rosenberg 1999 provides an impressive review of the potential structural complexities of hypertext, particularly of the literary kind.
2.3.2 Items vs. connections
Recall our working definition of hypertext (slightly abridged):
• hypertext is an interface to interconnected items
• the items are of information, textual or pictorial
• the interface lets the user follow any connection
• the interface records the connections followed, and lets the user relive them
Items, connections: cleary the two basic building materials of hypertext. These two materials are—we submit—at a tension, which constitutes a foundational aporia of hypertext.
Theoretically, any large enough item may be transformed—shattered—into, two items plus a connection between them. Conversely, any two items with the proper connection natural language is often taken as the prototype of language, and, accordingly, linguistics as the mother science of semiotics. We also note that the term media in hypermedia denotes something often also called content—which concept is, in a way, the very opposite of media. Probably semiots are more keen on the sensibleness of their terminology than marketeers—as it should. 6Collectivelly known in the learned lusophony as discurso gnosiol´ogico, cf. Belo 1991. 2.3. FORMS OF HYPERTEXT 21 between them may coalesce into a single item. Indeed, this observation is at the basis of the Shattered Documents approach of this thesis.
The Nelsonian concept of transclusion also blurrs the distinction between item and connection, cf. Nelson 2007b.
This aporia is perhaps better understood within an historical perspective. The current times are the age of connections. Little or null attention is given to the nature or construction of items. This focus on connections has been a constant since Bush 1945:
“The process of tying two items together is the important thing.” (Bush 1945)
Incidentally, note that this focus corresponds to the hyper part of the many hyper... words, which would in fact explain their proliferation.
Bush 1945 describes the memex, an hypothetical machine for memory extension, based on a network of microfilms and on an user interface equiped with means to explore this network. Indeed, the memex crucial attribute is, in the author’s own words,
“associative indexing, a provision whereby any item may be caused at will to select immediately and automatically another. This is the essential feature of the memex. The process of tying two items together is the important thing.” (Bush 1945)
Bush 1945 is consistently touted in the current literature (e.g. Mancini 2004) as the seed of hypertext thinking. But we have found descriptions of similar conceptions predating Bush 1945.
Earlier in the 20th century, there was a documentalist movement concerned with using the information and communication technology of the day—radio, x-rays, cinema, microfilm—to improve the global access to global knowledge. Otlet 1934, Otlet 1935 envision the convergence of such technology into an Office of Documentation, or Mundaneum, to form “a mechanical, collective brain... an exodermic appendage to the brain... a substratum of memory... an external mechanism and instrument of the mind” (Otlet 1934, Otlet 1935, apud Rayward 1999).
Note the obstinated use of brain metaphors. Clearly this line of thinking is already inscribed in a tradition of intelect augmentation7 by artificial means.
7Let us use the term introduced by Engelbart 1963, advantageously more general than the memory extension of Bush 1945, assuming that this is what the word memex was an abbreviation of. 22 CHAPTER 2. HYPERTEXT
“Intellectual power, like physical power, can be amplified. Let no one say that it cannot be done, for the gene-patterns do it every time they form a brain that grows up to be something better than the gene-pattern could have specified in detail. What is new is that we can now do it synthetically, consciously, deliberately.” (Ashbi 1956)
Returning to Otlet and the 1930’s: he proposes the organization and transmission of knowledge in a global network. The Mundaneum, conceived as the centre of the network, would be replicated at several levels. There would be a Mundaneum in every country, in every city, and, finally, every person could have their own technologically sophisticated office, called Studium Mundaneum, in which they could access the repertoires. Which would clearly impact the professional, personal, familiar and social life of every individual (Santos 2006, p. 97, quoting Rayward 2003, p.7). Clearly, Otlet envisioned no less than the Internet of today.
Also, the documentalists sustained the Monographic Principle as a basis for the Office of Documentation.
2.3.3 The Monographic Principle
The Monographic Principle takes form in the recording of single pieces of information onto standardized cards, with larger chunks of information recorded on separate sheets. These cards could be managed, copied, combined, in order to search, form, present information. (Otlet 1918, apud Rayward 1994). The approximation to the nodes of hypertext, specially in the Shattered Documents approach, is compelling. As Santos 2006 puts it: “Otlet reformats the document by fragmenting it, and reorganizes content, generating new informational wholes. The informational objects created by the Monographic Principle approximate today’s databases and hypertextual objets.” (p. 90).
Nelson would coin the word hypertext. Otlet created none other than the term: documentation!(Otlet 1918)
The monographic units, the fragments, were reorganized on the basis of classification systems e.g. Otlet’s own Universal Decimal Classification. The units were organized in sequences assigned to a certain classification. The UDC and other classificatory systems allow for multiple classification of the same item, so an item could participate in multiple sequences. So, structurally, these sequences are like the trails of the memex 2.3. FORMS OF HYPERTEXT 23
Bush 1945. Conceptually, however, they are a world apart.
The difference is between positivism (documentalists) and deconstructivism (memex). The documentalists assumed the possibility of a man-made, predefined classification system capable of organizing all knowledge, whereas memex hypertext is created, organized at will by the user, with or without resorting to any predefined classifica- tory system. Indeed Bush 1945 discusses the inadequacy of classification systems for hypertext. So, along with the general passing of positivism, eventually classification systems were abandoned as a means of connecting the items, and replaced with free linking by the usors.
Unfortunately, the baby was thrown with the bath water. Since Bush 1945 the focus has shifted towards the connectionist part of hypertext, and little or null thought has been given to the items connected. More often than not, the items are assumed as given, integral parts of the system: the pages. We have already noted in several parts on this thesis how this assumption lacks theoretic foundation, and how the respective design creates interface faults, for example long pages requiring scrolling, which extra cognitive effort distracts the user from their main quest. Pages must be shatterable. So, our approach of Shattered Documents may be seen as a return to the good part of Otlet’s documentalist theory.
In sum, Otlet created the monographic principle, but no hyperlinking. Bush envisioned hyperlinking, but sans monographic principle. And to this day the monographic principle as been dormant to say the least. Whether on a path of Hegelian dialetics or not, on this thesis we do try to make a small contribution to finally unite the two things—free connections, free items—in a long due synthesis.
2.3.3.1 More words: microtext, macrotext
We have seen how the notion of hypertext is often blurred with the notion of global access to global knowledge. In the current Word Wild Web the two aspects seem inextricable.
A certain branch of (modern) hypertext literature uses the terms microtext, macrotext, to refer to the local vs. global range of hypertext (Keep et al. 1995). The Web is a macrotext. The memex is a microtext.
A single Office of Documentation is local, but Otlet’s envisioned a network of such Offices spanning the World, and each replicating the collection of World’s knowledge. 24 CHAPTER 2. HYPERTEXT
Clearly a world of Internet caf´es.Macrotext.
Macrotext was the focus of another hypertext visionary—seldom recognized as such, albeit a very famous author: H. G. Wells. Aparently since 1902 (apud Rayward 1999), Wells has exposed a vision of a world brain, and a global encyclopedia, created by means of technology—notably the same technology and even the same methods selected by the documentalists (Wells 1938). Aparently Wells and Otlet met in the late 1930’s ou early 1940’s, cf. Rayward 1999.
(At this point, the science fiction connoisseur will probably evoke the Encyclopaedia Galactica or the Hitchhicker’s Guide to The Galaxy. We note in passing that neither of these devices seems to be a macrotext or even a microtext.)
Aparently H. G. Wells has also foretold the screen, or monitor, as we know it today, on his science fiction novels (apud active literature cited above).
2.3.4 Summary of hypertext history
For a summary of hypertext history see table 2.1.
Table 2.1: Summary of hypertext history Early 1900’s Otlet, Wells, envision the cyberfuture Otlet coins the term documentation 1945 Bush designs the Memex 1960 Nelson designs Xanadu 1965 Nelson coins the term hypertext Engelbart creates HyperCards (and the mouse) 1980 Berners-Lee creates HTML, a limited form of hypertext 1990 The Web develops within this initial philosophy 1995 The Web grows enourmously—lost in hyperspace 2000 Web 2.0, social networks—lost in MySpace
2.3.5 The document dogma
Circa 56% of the text on the Web is in the form of traditional documents stored in PDF format. 2.3. FORMS OF HYPERTEXT 25
This reflects a mentality, which we may call the document dogma, that the traditional document is the preferred, or somehow best, format to represent information. Either from the reader or the author point of view—the result is the same.
But it has been shown that hypertext—at least a moderate form thereof—is actually better for online reading than the imitation of paper (cf. Nielsen 1997).
However, traditional documents are hard to decompose into the independent units required for full-fledged hypertext, because traditional documents contain elements of—cross-unit—sequentiality or structure that give meaning to each unit involved.
Also, in the literature—and in designed or constructed systems—the nodes are always given. The various theories, systems, designs, focus on the links, on different ways of connecting the nodes. They never study, model, much less challenge, the internal construction of nodes. This assumption is probably in line with the document dogma mentality.
2.3.6 Information search—the impossible that is done
A search for information is a theoretical conundrum, because it aims at finding what is unknown; and therefore—theoretically—the search cannot be expressed, stated, in the terms of the sought solution. But as we all very well know, search for information happens all the time. This is so because the need for information is real, and must be solved, even if imperfectly. (The difference between theory and practice is that, in theory, there is no difference, but in practice, there is.) Some encompassing works of library science, documentation, and information retrieval, e.g. Foskett 1996, correctly recognize this effort of finding the unkown as the fundamental problem of information retrieval.
This impossibly perfect concept of information search entails the equally impossibly perfect, fixed concept of information need: “a constantly changing, inaccessible phe- nomena present only in the mind of the searching agent; a combination of ideas such as what the target information might look like, where it might be found, or how one might go about tracking it down—with the words look, where, and how used in a most general sense.” (Campbell 2000)
A single query of (classical) information retrieval subsumes a fixed information need. The real, dynamic, changing information need is often realised as a series of queries, with each subsequent one said to be a refinement of the previous. This dynamics is 26 CHAPTER 2. HYPERTEXT also the essence of relevance feedback techniques.
Hypertextual navigation (as well as the ostensive search model of Campbell 2000, which is essentially hypertext), as a form of information search, represents a slight departure from the classical query model. With hypertext, the user is presented, at each time, with predefined paths to follows.
In practice, users of the Web, and of hypertext, alternate between querying and nav- igation, mostly according to the circumstances. The integration of these two distinct search methods in a single interface has been a kind of holy grail for some researchers e.g. Olson & Chi 2003. In this thesis we do not pursue this type of integration. We simply focus on navigation.
2.3.7 Aporias of adaptation
2.3.7.1 Dictatorship of the majority
Of the two main kinds of adaptation—personal, colective—only the latter construes aporias.
Adaptation towards the colective mind brings the problem of dictatorship of the majority.
We assume that the collective brain is nevertheless useful to the individual in a large enough class of situations. There is plenty of (soft) evidence of this fact—the epitome being the very successful ranking of Google results based on PageRank, an algorithm that integrates data from many usors (and therefore is a kind of collective brain).
2.3.7.2 Unwarrant self-reinforcement
The essential design of recommender systems causes the following problem.
In a recommender system, the user is presented with recommendations. It is likely that the user follows the recommendations—or else the recommender system would be useless. But following a recommendation will strengthen the presence of the recommended item itself in the recommendation set. This veritable self reinforcement of the identity of the recommended set, i.e. of its set of items, has the vicious effect of keeping other possibly interesting items out of the set—and hence outside the user’s view. 2.3. FORMS OF HYPERTEXT 27
So far the only patch to this problem is to bypass the recommendation model proper by including random items in the recommendation set. Formally this represents a degradation of the result, and an artificial injection of noise into the system—clearly not a satisfactory situation (and hence an aporia).
(The identification of this problem, as well as the suggested patch, are, to the best of our knowledge, original contributions of this thesis. If I remember correctly, problem and patch were firstly proposed by myself in a meeting of a past European research project (Mladeni´c & Lavraˆc2003). Albeit discussion has ensued, it has remained un- published. However the problem is so evident, that we cannot but wonder how it has been treated by the myriad recommender systems out there in the commercial Web e.g. amazon.com.)
2.3.8 Minor issues
The aporias in this section seem only distantly related, if at all, to our main hypotheses. But they are too related to hypertext in general, or too culturally widespread, or just too interesting, not to be addressed in this thesis.
2.3.8.1 Anchoring aporias
On commonplace hypertext, links are anchored in the text it self, identified by a different typography. This design poses a number of problems.
When the term under anchoring appears multiple times on the same page, as often is the case, the design decision is not clear of which occurrence of the term to select as anchor. The first? All? Some subset? Which one? For example, SWI-Prolog designers chose to select all occurrences, cf. figure 2.6.
An alternate design, which dissolves this aporia, is to represent the links as separate items from the text, perhaps as buttons situated to the right margin of the page, and ranked by order of appearance on the text, as illustrated on figure 2.7.
As another, independent, example of this alternate design for help text, consider the interface depicted on figure 2.8. Note how the help items are strictly contained in the right panel—and not as hyperlinks anchored on the main text (the remainder panels), which would create visual clutter, and incur in the anchoring problems discussed.
Another problem of in-text anchoring is when the anchor is a compound term (i.e. 28 CHAPTER 2. HYPERTEXT
Figure 2.6: SWI-Prolog help for catch/3. Multiple anchors design. (Anchors are set in a green bold face.)
Figure 2.7: Alternative to the design in figure 2.6 2.3. FORMS OF HYPERTEXT 29
Figure 2.8: The web interface of the network router Linksys WRT54G. 30 CHAPTER 2. HYPERTEXT a multiword term) which comprises one or more subterms which are also potential anchors. This occurs extremely often in dictionary or encyclopaedic hypertext e.g. the Wikipedia. In-text anchoring simply cannot solve this problem. Consider, for example, the Dictionary definition of HTML, depicted in the centre of figure 2.9. The definition contains the compound terms Hypertext Markup Language and World Wide Web. These two terms encompass the eight terms entouring the picture, all entries in the Dictionary. However, the in-text anchoring design has made some terms not directly reacheable. The lines connecting the first letter of a word in the definition text represent clickable connections (remember that in Dictionary every word is clickable); the connections are to the indicated dictionary entries. Note that the entries markup language8 and world are disconnected. Note how the designers chose a special behaviour for the word World, namely selecting it as an anchor for the compound World Wide Web—disconnecting world in the process. And, of course, unsignalled different linking behaviours introduce unpredictability in the interface.
Figure 2.9: How compound terms are (mis)treated in Dictionary. markup language
hypertext markup language
World Wide Web web
world wide
Yet another problem, albeit less severe, of in-text, typographical, anchoring is when the anchor is not text at all, but e.g. a picture or an icon. In this case a typographical difference is simply not possible. But this problem only affects the typographical approach to anchoring, and can be solved relatively easily in a number of manners.
8The term markup language is meaning number 3 the entry for markup, requiring scanning, and normally scrolling, on this entry in order to be found; a direct link would avert this cognitive effort. All other terms are main entries. 2.3. FORMS OF HYPERTEXT 31
2.3.8.2 Data deficit
Invariably the access data, or HTTP log, is not rich or acurate enough for the expectations of the adaptation model.
A typical subproblem is the derivation of a session i.e. the identification of the accesses (clicks) that constitute a session by each user. Normally this can only be acomplished by the 30 minute rule—an heuristic rule based on the assumption that if a user is quiet for 30 minutes then their last request terminated a session.9 This rule yields results of varying degree—depending on the context or information domain—but that can never be 100% right.
Another subproblem is the identification of users, for cross-session derivations, user model construction, and of course personalisation. Only registered users can be rightly identified—but user registration impinges somewhat upon the Null User Effort principle.
Another problem is that the Back button serves distinct, even opposite, purposes:10
(1) The user follows a link that appears to be the oracle11; after going there, the user realises that it is not; the user pushes the back button simply to retreat from this false oracle, back to a point known to be of greater interest
(2) Same context as above, but Back button pushed simply out of panic. The Back button as a Panic button. The user is lost. They do not press the button to return to a Emknown place. They simply want to get out of there.
(3) The user has followed a link that appears to be part of the oracle—and it is. Then the user uses the Back button simply to continue exploring the web space from a more confortable place, probably a hub. etc.
Note item pairs (1)–(3) and (2)–(3) are pairs of opposites.
9In an concrete experience this value has been fine-tuned to 28 minutes. 10This problem is distinct from the Back button aporia described in a dedicated section. This problem is about interpreting Back button hits for adaptive purposes. The aporia is about the user expectations towards the interface. 11On this thesis oracle means roughly what the user is looking for. For the specialized meaning of this and other terms on this thesis see Annex A “Glossary” and parts referenced therein. 32 CHAPTER 2. HYPERTEXT
Note that access logs typically do not even contain Back button information. They simply stack the accessed pages. For example, for the session in figure 2.10(1), the log would simply record the sequence A-B-C-B-D-E. This is so in part because HTTP12 was designed as a stateless protocol. So, typical access logs, produced by the web server (e.g. Apache), are too low level for an adaptive system relying on user action information. There is a semantic mismatch. An adaptive system should maintain its own user action log. Several designs are proposed on this thesis.
2.3.8.3 Aporias of renoding
The concept of renoding, introduced in this thesis, carries an aporia of its own: if the page as we knew it ceased to exist, what else becomes the nature of links between pages?
In this thesis, and specifically in the shattered documents approach, this aporia is dealt with by introducing the concept of central noogramicle. Each page has a central noogramicle, which acts as representative of the page for linking across pages. The actual links are between noogramicles. This can be reinterpreted as connection between pages when the noogramicles are central.
2.3.8.4 Landmarks
Already well into the course of our investigation we have realized that Raskin 2000 is probably right in his observation that landmarks are a major help in navigation. Users looking for a certain item A, which they have seem before next to a certain item B, might find it easier to relocate A via B, because, for example, B is in sight (and A is not).
Landmarks and adaptive hypertext are irreducible. Landmarks require a stable, fixed territory which to mark and be inscribed upon. Adaptive hypertext is, by definition, a changing territory.
12HTTP = HyperText Transport Protocol, the communication protocol for WWW pages, cf. W3C. 2.3. FORMS OF HYPERTEXT 33
2.3.8.5 The Back button aporia
The Back button concept contains an aporia of its own, namely an unsolvable tension between its expected behaviour being chronological or not, cf. figures 2.10 and 2.11.
Figure 2.10: The Back button aporia. In certain contexts the user expects a chronological behaviour.
(1) B A C (back)
D E here user wants to go back to C
(2) user presses Back three times, (3) but A is where he gets: expecting to travel thus: B B A C A C
D E D E
Figure 2.11: The Back button aporia. Why the Back button should not behave chronologically.
(1) B A C now user wants to get back to A
(2) user presses Back two times, (3) if Back button behaved chronologically, being delivered to where expected pernicious effect would take place B B A C A C 34 CHAPTER 2. HYPERTEXT
2.3.8.6 Lost in hyperspace
The phrase lost in hyperspace has been echoed in the literature, with two meanings:
(1) lost in the hypertext structure, because of the hypertextual nature
(2) lost in the enourmous sea of information of the Web and modern culture in general
We downplay this purported aporia.
We believe that problem (1) has been solved with the Back Button and akin devices e.g. the Site Map. Golovchinsky 2002 provides a discussion of the Back Button, including advanced variants. The Site Map is commonplace in the Web today. Other navigational aids include: the fish eye, prototyped on Campbell 2000; the ZOOM interface, reported on Raskin 2000; the Map View, commonplace in videogames; the page trail, proposed in the current thesis.
Problem (2), in the Web, is partially solved with search engines e.g. Google. A better metaphor would be: drown in hyperspace. Or, with the advent of the so called Web 2.0, and its social networks, problem (2) becomes more like lost in MySpace—which might happen well before one reaches one million “friends”: according to recent results from psycho-social studies, 150 is the approximate number of direct contacts that a single person is able to manage (Dunbar 1993).13
Problem (2) in general equates the well known information overload problem. We shall observe that, contrarily to a common conception, this problem is not specific of our times, say of TV and Web times, but it has emerged, and has been analysed, and fought, since as early as the begining of the 20th century at the least. It is the very problem addressed by Otlet and the documentalists, as already discussed (sections 2.3.2, 2.3.3, Otlet 1934, Otlet 1935, Rayward 1999).
2.3.8.7 Design by tekkies
The design of web hypertext is extremely technology driven. HTML, JavaScript, Flash, whatnot, set the boundaries of what is done.
13MySpace, FaceBook, LinkedIn, Ecademy, HiFive, Plaxo, OpenSocial, SoundClick, Ning, Indaba... Albeit the social networking phenomenon is here to stay, and it is clearly related to collective adaptiveness in potential, in this thesis we had to stay away from the issue, because it is too large an issue on its own, and clearly the dust has not settled on it yet. 2.3. FORMS OF HYPERTEXT 35
Technologues are not only in charge of designing the underlying technology—much too often they impinge on the area of web design itself, of the use of said technology. Most disastrous designs that plague the web probably stem from this fact.
“The purpose of hypertext was always to make up for the deficiencies of paper. Paper cannot easily show connections, has very limited space, and forces an inflexible rectangular arrangement. Hypertext, the generalization of writing, potentially offers many forms of interconnection and presenta- tion beyond what paper allows. But so far the mechanisms, not the users, have been put in control.” (Nelson & Smith 2007)
Nelson also affirmed that computer interfaces should be designed not by tecnologues— as the Web was—but by filmmakers. He points out a lesson from history that the first films—made by the technologues of the day—were very poor in design or art. Recalling the prototype of first film, La Sortie des usines Lumi`ere (1895), we cannot but agree.
See also: Nelson 1999 and references thereof; Nielsen 1997 and other articles by this columnist on usability; Lebedev 2007 for design insights (and insightful design).
2.3.8.8 The computer as hypertext
One could argue that any clickable item on the computer constitutes hypertext. Menu items, file icons, window controls... the WIMP14 as hypertext. The general pattern of the definitions seen so far certainly apply. The only provision necessary is that the nodes (pages) may be also actions, operations of the computer. Even the Back button requirement is met:
• it is right there on certain WIMP interfaces, notably file navigators like Mac OS’s Finder, Windows Explorer...
• in a limited number of contexts, the WIMP has it under other forms, namely the Escape key (to regress from a submenu or window), or the Undo operation (often bound to the Ctrl-Z key combination).
So there a bit of hypertext in the WIMP after all—which is good.
14WIMP = Windows, Icons, Menus, Poiting device (mouse): the interface quartet on most if not all computers today. 36 CHAPTER 2. HYPERTEXT
Furthermore, there are alternate arrangements to the WIMP, including operating systems based on hypertext (N¨urnberg et al. 1996), and whole computer designs biased towards—or strictly for—information recording and retrieval by the user (Raskin 2000).
More commonly, we shall observe that many of today’s interfaces are indeed realised as hypertext of the web kind, either because they rest upon a web service architecture, or simply for convenience, or because the interfaced object is already a network server or device, e.g. mail servers, network routers (figure 2.8).
2.3.8.9 The logical structure myth
“LaTeX is different from other typesetting systems in that you just have to tell it the logical and semantical structure of a text” (Oetiker et al. 2005, emphasis mine). The same has been said of HTML.
Such assertions are misleading to say the least. In any case they represent an enormous abuse of the concepts of logic and semantics, because Latex and HTML commands are completely document-oriented: chapter, section, heading, paragraph, itemize, unordered list, enumerate, ordered list, table, etc.—all are elements of form. As are the styles of WYSIWYG15 word processors like Microsoft Word.
A truly logical, semantical, abstract, rational, system must be oriented towards rhetor- ical categories or rely on text generation devices. Indeed, so far, only text generation research seems to have produced such logical text devices, e.g. Power et al. 2003. And the art is still confined to the laboratory.
2.4 Structure of documents
2.4.1 Traditional document structure
Traditional documents portray a sequential organization. Long or complex documents also feature organizational devices that are hierarchical, or cross referential, or both.
The sequential nature is either imposed, to a lesser extent, by the physical form of the codex, or, to a greater extent, by the physical form of scroll—which, as we have seen in chapter 2, is a millenia-old book form, resurrected, to our misfortune, in the
15WYSIWYG = What You See Is What You Get. 2.4. STRUCTURE OF DOCUMENTS 37
current computer age. The sequential nature of documents is also associated with the sequential nature of narrative discourse.
Hierarchical devices consist mainly in the hierarchical containment structure of chap- ters and sections. Finally, cross referential devices include: references in one place to another place e.g. “for ... see section X”, “see also section Y”; back of the book indexes e.g. alphabetical index; the Table of Contents.
As for the substance itself of documents, we observe that is consists of figures, formulae, listings, tables and, of course, free text, i.e. text that is situated outside any figure, formula, listing or table. Text is the usual substance of headings and paragraphs.
For the purpose of the present thesis, it is clear that text may be analysed into constituents like headings, paragraphs, sentences—but no further into its smaller, or physical, units like words and characters. In the same vein, figures, formulae, listings and tables are unlikely to be analysed into their constituents. Just to clarify: we do not analyse documents at the level of characters, lines or pixels.
Finally, we observe that the items that constitute a document fall into two categories:
Propositional, or narrative units: figures, formulae, footnotes, listings, paragraphs, sentences, tables. These are units that, all by themselves, form a proposition, i.e. they assert something, or tell a story. These are the units more likely to ultimately satisfy an information need; in Shattered Document terms, these units serve as oracles.
Indicative, or non-narrative units: captions, headings, titles. These elements only indicate, or name, other, normally propositional, parts of the document. They do not make an assertion, or proposition, by themselves. These units are more likely to help in navigation, to guide the search for the oracles, than to serve as oracles themselves.
2.4.2 Standard hypertextualization
The various components of a traditional document are amenable to hypertextualiza- tion in various degrees. Figure 2.12 depicts the standard hypertextual edition of the Ada Reference Manual, or ARM. The representation, albeit stylised, honors the actual data, in the numbering and size of the sections, and the links thereof. Each section is a node, or page—an integral HTML file. 38 CHAPTER 2. HYPERTEXT
The pages are connected by Next and Previous links. Four sequent nodes in this sequence are represented. The links are symbolized by the arrows in figure 2.12, and designed on the interface as buttons located at the top and at the bottom of each page, cf. figure 2.13 which captures the real rendition of the first page in the sequence.
This sequential linkage, with Previous and Next buttons, is ubiquitous on hypertex- tualized large documents. The hierarchical and cross-referential linkage, on the other hand, is found in varied forms.
A Table of Contents (or Site Map) may or may not be present. Naturally, such device should provide direct links to the listed sections of the document. Alternatively, or cummulatively, some or all of the pages or sections may contain a sub-Table-of- Contents, providing such direct access to the their subsections.
An Up or Top button (or both) also may or may not be present in the current page or section, to allow moving upwards in the hierarchy. The Top button may lead to the Table of Contents—probably the most sensible design—or to other high up pages like the document cover, the start of the first top level section, or the start of the current top level section (occasionaly a Top of Section button is included, dedicated to the latter case).
Our example hypertextualised document, the ARM, contains a Table of Contents, and every page includes a button to it (fig. 2.13). Also one click away are: an alphabetical index; a text search facility (only works online); a list of related documents (with links); the cover page (enigmatically called Legal Information). Each of these items is also, like sections, a single node in the hypertext structure.
The text of sections contains cross-references to other sections. These links are to the sections as a whole. Naturally each item of the Table of Contents also links to the respective section as a whole.
Although each page is an entire section, the ARM has the notion of paragraph, properly implemented in the hypertext as a page fragment i.e. an element within an HTML document. The paragraphs are numbered within each section, starting at 1. Whereas the distal content of each Table of Contens item, and of each intext cross-reference, is a whole section, or page, the distal content of each entry in the alphabetical index is a numbered paragraph. 2.4. STRUCTURE OF DOCUMENTS 39
Figure 2.12: Standard hypertextualization of the sequential structure. 1. 1.1 1.1.1 1.1.2
Figure 2.13: Real look of the first page in figure 2.12. 40 CHAPTER 2. HYPERTEXT 2.5 Learning systems
Adaptation, in an adaptive hypertext system, consists in automatically changing the presentation of information and the overall link structure towards the needs of the user (Bra et al. 1999). The benefits over a non adaptive system include:
Reduced authoring effort. “Adaptive Hypermedia Systems (AHS) make it possi- ble to deliver ”personalized” views or versions of a hypermedia document (or hyperdocument for short) without requiring any kind of programming by the author(s).” (Bra et al. 1999)
Improved usability. Hopefully, “information can be found faster or can be better comprehended when adaptive techniques are used; the use of adaptive hyperme- dia (presentation and/or navigation) is beneficial, for comfort or performance.” (Bra et al. 1999)
Brusilovsky & Mill´an2007 call adaptive effect to the effect that each user receives a different view. That term is misleading because adaptation can occur sans a different view per user—but towards the document as one, or towards a collective or the whole of users, as we shall see. Perkowitz & Etzioni 1999 acknowledge that fact and, consistently, call customization to the specific view per user effect. Another preferable word used in the literature at large is personalization.
The research on adaptive hypertext is scattered over a quite large number of areas including such ones as User Modelling,16 Intelligent User Interfaces,17 Adaptive User Interfaces,18 Intelligent Tutoring Systems,19 Human-Computer Interaction,20 Adaptive Hypermedia,21 Learning Systems,22 Hypermedia Operating Systems,23 Cybernetics.24
16Brusilovsky & Mill´an2007, Herder 2003, Perkowitz & Etzioni 1999, Sos- novsky & Brusilovsky 2005.
17Kov´acs & Ueno 2006.
18Paramythis et al. 2001, Perkowitz & Etzioni 2000. 19Czarkowski 2006 and references therein.
20Card et al. 1983, Cockburn et al. 2002, Heylighen 1997, Olson & Chi 2003, Raskin 2000.
21Alves et al. 2004, Bra et al. 1999.
22Ahn et al. 2005, Sosnovsky & Brusilovsky 2005, KnowledgeSea.
23N¨urnberg et al. 1996.
24Ashbi 1956, Bollen 2001, Bollen & Heylighen 1998, Heylighen 1997, Heylighen & Joslyn 2001, Principia Cybernetica Project. 2.5. LEARNING SYSTEMS 41
The latter two areas are probably the most island-like (each on its own). It is revealing that the bibliography of N¨urnberg et al. 1996, a prime representative of the Hypermedia Operating Systems area, is totally disjoint with the bibliographies of the other areas. Regarding the efforts from Cybernetics, e.g. Heylighen 1997, Bollen & Heylighen 1998, Heylighen & Joslyn 2001, Bollen 2001, they are mostly unbeknownst to the others as well—whereas the Cyberneticians themselves seem knowledgeable of those. In this thesis we do not situate ourselves strictly in any Capital Letter Research Area. We would gladly do so regarding Hypertext Science—if such one existed.
In fact, a truly general theory of adaptive hypertext, spanning all areas, does not exist yet—altough the field is fraught with conceptual frameworks—also called reference models, general models, architectures, etc. For example, the Dexter reference model is popular in the Adaptive Hypermedia and Learning Systems areas. Partially dependent on the Dexter reference model, a general theory of modular evaluation has been discussed (Paramythis et al. 2001, Herder 2003, Sosnovsky & Brusilovsky 2005).
2.5.1 Adaptive hypertext techniques
The opposition presentation/navigation discussed on Bra et al. 1999, and on references thereon, is also a popular ontological basis for organizing the world of adaptive hy- pertext models, or techniques. Most items and descriptions below are taken from Bra et al. 1999, but were reorganized, edited, corrected, or extended.
Presentation techniques include:
Conditional inclusion of fragments. Only fragments that are recommended for a user are displayed.
Stretchtext. For each information fragment there is a (short) visible place holder. The system determines which fragments should be stretched (i.e. shown) and which fragments should be shrunk (i.e. only the place holder is shown). This decision only determines the initial presentation of the fragment. The user may stretch or shrink fragments through mouse clicks.
Graying. Fragments that are not recommended for a user are grayed out (instead of excluded or shrunk).
Explanation variants. The same information can be presented in different ways. This can be done within a page or through guidance towards different page 42 CHAPTER 2. HYPERTEXT
variants. (In the latter case the method becomes adaptive navigation support rather than adaptive presentation).
Reordering. The order in which information items are presented may have to be al- tered. For instance, some users may prefer to see an example before a definition, while others prefer it the other way around. On a page, fragments of information are typically sorted from most to least relevant, a method which is best known from information retrieval systems.
Other techniques. Frame-based; natural language generation; dynamic hypertext, combining querying and linking in order to provide personalized link structures; cf. Bodner & Chignell 2000.
Adaptive navigation techniques include the following. Mind that the word link might actually denote—and often does—an anchor, or else entail the existence of an anchor.
Relinking. The target of the link is changed.
Direct guidance. A ”next” or ”continue” (link) button.
Sorting of links. A list of links is sorted and presented from most relevant to least relevant.
Link annotation. Link anchors are presented differently depending on the relevance of the destination.
Link disabling. Inappropriate links are disabled.
Link removal. Recall that this is actually anchor removal. So, naturally—and as Bra et al. 1999 themselves acknowledge—in running text this is a big dont, as it simply ruins the text.
Map adaptation. Some hypermedia systems provide a graphical presentation of (part of) the link structure. Such maps can also be subject to adaptation.
We must note that the technique of relinking is notoriously lacking on Bra et al. 1999. Notoriously but not surprisingly, given the seeming absence of communication between research areas. Relinking is a staple of the Cybernetics approach to adaptive hypertext (Bollen 2000), which results are unbeknownst, apparently, to the Learning Systems and Adaptive Hypermedia dwellings of Bra et al. 1999. 2.5. LEARNING SYSTEMS 43
Bra et al. 1999 also list the technique of link hiding, whereby “links leading to inappro- priate or non-relevant information are hidden” (Bra et al. 1999). We do not understand the difference of this technique with link disabling, or link removal. If link hiding affects only the linking function, then link hiding equates link disabling. If it somehow affects the anchor, then it equates link removal.
Note that this classification of techniques depends on a tripartite classification of hypertext elements into pages, fragments and links (Bra et al. 1999). It is acknowledged that pages are made up of fragments. But the monographic principle is never at work here. Each page is given with its fixed set of fragments. Only the way each fragment (or link) is presented is subject to adaptation.
It is also interesting to analyse the parenthetical remark on the definition of the Explanation variants technique, adapted here for convenience:
“In the case of guidance towards different page variants, the method be- comes adaptive navigation support rather than adaptive presentation.” (Bra et al. 1999)
This shows that the authors acknowledge a possible frailty of their presentation/- navigation dichotomy as a classificatory device. This kind of conceptual classifica- tion attempts—and failures—pervades the literature. Perkowitz & Etzioni 1999 provide another system of classification of adaptive hypertext techniques—also dual, but traversal to the presentantion/navigation system:
“Adaptive sites may use customization or transformation. Customization is modifying a Web site to suit an individual user; the page returned by the site depends on the user model. A site may also perform online customization—customizing the pages served to the user as she browses. In this case, the site takes into account not only the user model, but also the pages just visited in choosing the next page to show the user. In contrast, transformation involves altering the site to make navigation easier for a large set of users. For example, a university Web site may be reorganized to support one “view” for faculty members and a distinct view for students. In addition, certain transformations may seek to improve the site for all visitors.” (Perkowitz & Etzioni 1999)
In our own efforts to either understand or design adaptative hypertext systems, we found ourselves repeatedly thinking in terms of the following concepts; and doing 44 CHAPTER 2. HYPERTEXT so more than in terms of—or else in complementarity to—the classificatory theories above.
The personal—collective opposition, or continuum. Does the adaptive input draw on personal (the user) or collective data (other users)? Of all users? Or of a subset (group) thereof? Which group or groups?
Adaptive input/output. What is the adaptive input specifically, i.e. what data is used to drive the adaptation process? What is the adaptive output specifically, i.e. what is the result of the adaptation process?
Interface. How does the user interface look like? How does it integrate with the adaptive input, output, and process?
Our own approach of Knowledge Atoms, like the KnowledgeSea, looks into the collec- tive side of adaptivity, as input. But, as output, we propose a more extreme kind of adaptivity whereby each page itself is built—at the time of request—from the shattered fragments of the document.
2.5.2 Learning Systems highlights
In our own survey of the variegated landscape of adaptive hypertext research, and particularly of the cluster around Learning Systems, two items have made a singular impression upon us: the overlay model, for its ubiquitousness, clarity, definiteness, simplicity, understandability, usefulness; and the system Knowledge Sea, for being the only learning system that draws on the collective mind, and in a clear way.
Most adaptive hypertext systems, particularly those of the Learning Systems and Intelligent Tutoring Systems denominations, are based on an overlay scheme, or model, whereby the content is overlaid with a measure of, the knowledge of the content, by the user. In other words, each content element is tagged with a measure of its being known by the user. In sum, the content model is overlaid with the user model. (Sosnovsky & Brusilovsky 2005)
Proceeding from this overlay measure, each system then applies an adaptation tech- nique to the presentation of the content.
Knowledge Sea (Ahn et al. 2005, KnowledgeSea) was the first system to exploit the collective mind, for adaptation. The authors call social search to this collective aspect 2.5. LEARNING SYSTEMS 45 of the system (Ahn et al. 2005). The name Knowledge Sea comes, most probably, from the visual aspect of the interface, made of tiles of different shades of watery blue, thus evoking the image of a wavy sea seen from above (figures 2.14, 2.15). Each tile represents a piece of content. The darkness indicates its popularity. Other gauges are available. The person icon denotes the user. Again, depth of blue indicates degree of use. Let the authors’ own word provide the details:
“Knowledge Sea is a Web-based social navigation support system. It organizes Web- based open and closed corpus C language teaching materials including online tutorials and lecture slides. In order to implement this mixed corpus based social navigation, Knowledge Sea uses a knowledge map of the domain—a two-dimensional table con- sisting of 64 cells. It is built by a self-organized map (SOM) algorithm. Semantically related keywords and documents were assigned for each cell. Contents of neighboring cells are semantically related. Background colors of the cells indicate the popularity of the cells. As more users click and visit a cell, the background color of the cell gets darker. When they click a cell, they can see a list of documents and can choose a document from the list. The same logic to represent popularities by color lightness is applied to the representation of documents inside each cell. Each item of the list provides two types of information, traffic and annotations. “Human-figure” icons and colors provide users with popularity information and “thumbs-up” or “thermometer” icon and colors provide users with annotation information. If a document is popular among the group where a user belongs to, the background color of the icon gets darker. The foreground color of the icon gets lighter if the user clicked the document fewer times than other group members. Just like popularity, darker background color of anannotation icon indicates there are a lot of annotations for the document. “Sticky- note,” “thumbs-up,” and “question-mark” icons indicate “General,” “Praise,” and “Question” annotations respectively. A red “thermometer” icon indicates the overall annotations are positive, and a blue icon indicates the overall annotations are negative. Therefore, users can navigate socially by referring to other users’ behavior and opinions by looking up these icons and colors provided by Knowledge Sea.” (Ahn et al. 2005)
Empirical evaluation of Knowledge Sea, reported on Ahn et al. 2005, has yielded an interesting result: the most relevant items were not the most popular, but the most annotated. 46 CHAPTER 2. HYPERTEXT
Figure 2.14: The Knowledge Sea interface (KnowledgeSea).
2.6 Summary
This long chapter presented the definition of hypertext, the terminology used in this thesis, and a rather philosophical offering about the aporias of hypertext. Next we will concretize our ideas into concepts that are scientifically manageable, either formally or in the computer. 2.6. SUMMARY 47
Figure 2.15: Detail of The Knowledge Sea interface (KnowledgeSea). 48 CHAPTER 2. HYPERTEXT Chapter 3
A new model for adaptive hypertext
Research is what I’m doing when I don’t know what I’m doing. Wernher von Braun
We present a new model for adaptive hypertext based on the concept of renoding: the instantaneous construction of pages from a ranked selection of noogramicles (the small parts of documents). We define this model with formal rigour.
3.1 Motivation
3.1.1 Information, not documents
We put forth the shattered document approach, with its postulate of noogramicles making up the documents, as an atomic theory of information needs. Our main premise is that users look for information, not documents. Whether users look for information or documents has been a long standing debate in information science, particularly in library studies (Otlet 1918, Foskett 1996).
In pre-computing times the issue obviously referred to physical, printed documents. With these being the only means of recording written information, the argument could
49 50 CHAPTER 3. A NEW MODEL FOR ADAPTIVE HYPERTEXT be made that users look for documents only inasmuch as they look for the information contained in the documents. That is, users know about the physical constraints, and plan their search accordingly. In fact there is ample evidence that this is indeed the case. Most consultation acts result in the selection of a part of a document, or in a combination of parts of one or more documents. The latter activity has even gained professional status and a name: clipping.
The counter-argument would be based on a conception of documents that are perfectly suited for each information need. That is, a document is a knowledge organization device. A well constructed document represents the consummate answer to a specific knowledge need. So the clipping activity is only necessary for non ideal cases, i.e. when the sought knowledge has not been the subject of such a document construction yet. However, the absolute number of non ideal cases is very large. And, specifically, large enough to justify any effort—such as the present thesis—of making it easier for the user to find the right information contained in the imperfect documents.
With the advent of computing, one would expect information to be recorded in a more maleable way than the document. This has in fact happened; two structures have become common: databases, hypertext.
The undeniable need for databases clearly reveals the atomic nature of a vast class of information needs. Virtually all database content is compound of atomic pieces of information, and queries are based on this microstructure. (The need for hypertext is much harder to analyse—we tried our best on the dedicated chapter 2.)
In conclusion, one can safely say that a very large number of information needs can be satisfied with small, atomic, parts of documents. 1
3.1.2 Guidelines for adaptive hypertext
This section is an evolution of a part of Alves & Jorge 2004. It is a set of top level requirements and design guidelines for adaptive hypertext systems: No Documents, No Scroll, No Intrusion,2 Full Control, Clear Labels, Full Incrementality, Best Practices,
1The school of critical thinking (Paul & Elder 2012) acknowledges the existence of problems that are monological. Such problems correspond to atomic information needs. 2Also called Null User Effort occasionally on this thesis or on related software artifacts. Historical reasons. 3.1. MOTIVATION 51
3 Pervasive Adaptivity. This is mainly a result of years of personal experience both as a user and creator of web systems, but these results are also supported by respectable authors, duly cited.
No Documents. Users look for information, not documents. (We are referring here to traditional, long documents: books, articles, etc. Not innovative, short items like cards, noogramicles, which might also be called documents in certain contexts.) Documents are the pre-digital medium of recording and communi- cating information. The document metaphor on the digital world is unadjusted (Coombs et al. 1987). The system must give the sought information right there on the screen, formatted for the screen. The online page is the screen, or the window. “The world of the screen could be anything at all, not just the imitation of paper” (Nelson 1999). See also No Scroll. Furthermore, documents are for reading, but “online users don’t read—they scan” (Nielsen 1997). Clearly here Nielsen is referring to long texts as (bad) containers of oracles. If a user is looking for a certain information, they rather do it on a scannable content, e.g. a short bulleted list, than on a long running text.
No Scroll. A page that requires scrolling is not a page, it is a document—and as such it is banned (see No Documents). Online pages should fit the user’s screen, or the window. Scrollable windows limit the view. Raskin 2000 has shown that scrolling had a impact on the reading effort greater than that of turning or invoking pages. A document requiring scrolling is actually an even more ancient information medium than the modern book, or codex: it is a scroll. So scrolling in the online world is a regression of more than 1000 years, to pre-codex age.
However, the findings of Raskin 2000 are in part derived from the difficulty of positioning the mouse pointer on the scroll bar—and therefore this part has disappeared with the mouse wheel. Also, it should be noted the recent trend of very long web pages in very popular sites such as Wikipedia or Facebook. Their popularity seem to challenge this
3Please keep in mind that the concept of best practices is distinct from that of common practices. Unfortunately, general discourse often confuses the two. 52 CHAPTER 3. A NEW MODEL FOR ADAPTIVE HYPERTEXT
principle (No Scroll). Although this is an ad populum argument, it should be taken in consideration. So we are not adamant in this principle. We are nevertheless still convinced scrolling is as least as costly as page turning or invoking (by clicking) in the context of document consultation—which is our object of study.
No Intrusion. The user should not be imposed any work other than his normal behaviour. For example, usability questionnaires are banned. Adaptive input should be restricted to the user normal activity. This also applies to authors. Regarding them, the baseline effort shall be that of the normal writing of texts including placement of non-text objects. The general baseline effort is that of non-adaptive systems. The user shall not be imposed any extra effort with respect to this baseline. Exception: Full Control may require extra effort for user levels above normal users and authors, e.g. configurators. Cf. usability studies e.g. Nielsen 1997. Deviation (hypothetical): some users may be willing to provide feedback. Most of this feedback is valid i.e. it is not intentionally boicoting. This set of users is representative of the whole. Therefore the system should provide a way to collect feedback, and use this feedback as valid knowledge.
Full Control. The system must allow the user to control every aspect of the adaptive behaviour of the system. In particular, it must allow the user to turn adaptivity off (and back on), to escape artificial stupidity (Raskin 2000).4 The means to exert this control must be readily available, right there alongside links or main options on the current page.5
Clear Labels. A link must be, at the source, or anchor, clear about its target, or distal content. That is, the interface must be so that users can effortlessly make a reasonably clear idea of the effect of activating the link—before activating it. A clear label approximates what Mancini 2004 calls a “predictable connection” (sec. 1.4.1 of Mancini 2004)
4Recently a concept of scrutability in adaptive hypertext has been proposed (Czarkowski 2006), which is essentially the same as this requirement of Full Control. 5And not buried under multiple layers of obscure menus, as is often the case, unfortunately. Recall Microsoft Excel or Word’s intelligence about capitalizing the first letter, or automatically numbering items—quickly, how do you turn it off? 3.1. MOTIVATION 53
Full Incrementality. Any operation on the system must be doable at any time, without requiring any kind of restart of the system, or stopping any functionality.
Best Practices. The system must be effective (do the job) and efficient (do it well). This is a general requirement of all engineered human interface systems, but we chose to state it explicitly because ever more often web systems are utterly unreliable or slow. Slowness, in particular, is a modern plague. It totally inverts the technology promise of making things easier and faster. Much too often we experience this: a person waiting for the technology, waiting for the computer to respond. Often the system is slow to the point that the corresponding manual procedure would be faster. Also, response latency induces interaction errors: repeated clicking on objects that should have reacted or disappeared, but that unnervingly have not done so, is a common phenomenon, which may cause malfunction.6 250 milliseconds. This should be the maximum response time of the interface to any user action.
According to Auber 2002 (who refers to Ware 2000 for details), 50 ms is the maximum delay between an action of the user and the displayed result on the screen for the user to believe in a causal link. Heer et al. 2003 (who refer to Card et al. 1983) indicate that this time can be relaxed to 100 ms. Raskin 2000 re- laxes further to 250 ms. Nielsen 1993 analyses the issue into the “three important limits”:
0.1 second is about the limit for having the user feel that the system is reacting instantaneously, meaning that no special feedback is necessary except to display the result. 1.0 second is about the limit for the user’s flow of thought to stay uninterrupted, even though the user will notice the delay. Nor- mally, no special feedback is necessary during delays of more than 0.1 but less than 1.0 second, but the user does lose the feeling of operating directly on the data. 10 seconds is about the limit for keeping the user’s attention focused on the dialogue. For longer delays, users will want to perform
6Slowness of web systems is often attributed to network latency. We have observed that often the slowness is due to poor engineering of the server or client, including pages too large (XML is too verbose), little or null cache, little or null asynchronous updating, poor database administration (lack of proper indexes, little or null query optimization), etc. In sum, not applying best practices. 54 CHAPTER 3. A NEW MODEL FOR ADAPTIVE HYPERTEXT
other tasks while waiting for the computer to finish, so they should be given feedback indicating when the computer expects to be done. Feedback during the delay is especially important if the response time is likely to be highly variable, since users will then not know what to expect. (Nielsen 1993)
Other best practices include:
• search facility a la Google, preferably deeply integrated with the adaptive model (a fine discussion of the difficulties of integrating different search models is found in Ferr´e & Ridoux 2004) • site map, but only for sites where such a facility makes sense
Pervasive adaptivity. No element of the system should be excluded a priori from adaptivity. Notably, pages should be no less subject to adaptivity than links. That is, relinking—yes; but also: renoding. However, note that this feature is yet to be proven useful, and is still confined to the laboratory—including the current work.
3.2 Model design
3.2.1 The Shattered Documents model
The Shattered Documents approach prescribes that documents be taken apart into their smallest constituents of meaning, or noogramicles. Naturally the noogramicles must connect with each other, in order to create, ultimately, a navigationable network, or hypertext.
We will look firstly—and mostly—at hypertext documents created from traditional documents, like our running example the ARM. We have observed three dimensions in the traditional structure of documents—sequence, hierarchy, cross-reference. We transcribe these dimensions into types of connection in the network.
We have also observed that documents are made up of small narrative and non- narrative units of meaning. Clearly, these are the units that become noogramicles in a Shattered Documents model. Therefore the Shattered Documents model of documents 3.2. MODEL DESIGN 55
Figure 3.1: Model of the same document in figure 2.12 but with the shattered document approach and the two types of connection Next (N) and Child (C). C C
N 1. 1.1 1.1.1 (simile) C NNC CN
N C N C
C N
C N
Figure 3.2: Page made up of document fragments. 1.
contracted views 1.1 (of the 2nd text parag. of sec. 1, heading of sec. 1.1 and 1st text parag. of sec. 1.1, resp.)
is a network structure of two types of node and three types of connection (a di-nodul- tri-nexial model).
Figure 3.1 shows how the same part of the ARM in figure 2.12 is modelled with the shattered document approach and the two types of connection Next and Child. Figure 3.2 shows a page made up of the first five noogramicles in the model; the constituent noogramicles of this page are selected using spreading activation from the first one in a manner detailed later.
We have designed this model with the following desiderata in mind:
• applicable to a vast range of existing documents 56 CHAPTER 3. A NEW MODEL FOR ADAPTIVE HYPERTEXT
• preservative of the original structure and content of legacy documents
• adequate for processing by the shattered documents approach, including spread- ing activation
• adequate to incremental authoring
To consolidate: a document is represented as a graph, or network datamodel of noogramicles interrelated by directed connections of three types—Next, Child, Refer— , as follows.
Next represents the linear order of paragraphs. Example: from a paragraph to its immediate successor. Note that, by the extended paragraph definition, Next also connects from a section heading to the first classic paragraph of the section, and from the last classic paragraph of a section to the heading of the next section.
Child represents the immediate subordinate relationship between paragraphs. Ex- amples: from a section to each of its subsections; from a paragraph introducing an enumeration (e.g. a bulleted list) to each item of the enumeration; possibly, even from a terminal section (i.e. a section without subsections) to each of its paragraphs
Refer represents other reference relationships. Examples: from a paragraph to a foot- note; from a paragraph to another paragraph or section (e.g. the so-called cross references, and see also references); from a index entry to its target paragraph or section; from a TOC entry to the corresponding section.
In the Child and Refer relationships, a section is represented by its first paragraph, normally a heading.
Original references anchored in sub-paragraph units (e.g. words) are represented as references anchored on the paragraph as a whole.
3.2.2 Adaptive information, and author as first reader
As we are targeting an adaptive system, we need a way to represent the corresponding information. The adaptative part of our model is compound of two main itens: 3.2. MODEL DESIGN 57
Pages. What the reader sees. A page is assembled from a small number of noogram- icles, in a manner detailed later. Naturally the user may navigate to another page. Pages are the adaptive output of shattered documents.
Travels. The navigation steps that usors (authors and readers) make in the document. Each travel is recorded, and used in adapting the construction of pages. Travels are the adaptive input of shattered documents.
The adaptative process integrates the two items, by assembling pages based on travel information. The main idea is to select the noogramicles that are most connected to the current one.
So, we must add the connection type Travel to the trio Next, Child, Refer already explained. Therefore, so far our document model is a network of noogramicles with four types of connection: Child, Next, Refer, Travel.
Finally, we interpret each of Child, Next, Refer, as Travel. This reinterpretation of the traditional document structure connection types Child, Next, Refer allows us to solve the “cold start problem”, and simplifies immenselly the process of spreading activation, as we shall see. This step is justified mainly because, if you look at it, the connections Child, Next, Refer are indeed the travels that the author intended the reader to make in the first place. Next is directly so. Child, Refer are contigently so—they are the paths laid out by the author for the reader to cross, wanting. Or, Child, Refer carry a rhetorical value—which amounts to the same effect (a contigent choice by the reader to follow).
3.2.3 Interface design
Our system is an interface into a large document. The interface unit is the page. Pages are accessed one at a time, normally. The document as a whole is partitioned, shat- tered, into fragments smaller than the page. These fragments are called noogramicles, for small pieces of knowledge.7
Each noogramicle has two renderings, or views: expanded, contracted. The expanded view is the noogramicle itself, normally. The contracted view is a clear label of the
7Noogramicles are also called knowledge atoms, or simply atoms, on other parts of this thesis, specially on software artifacts. Other words that were also, at a time, tried, and possibly used, include fragment, extended paragraph, and even texticle (for small text—this one was quickly dismissed though...) 58 CHAPTER 3. A NEW MODEL FOR ADAPTIVE HYPERTEXT noogramicle. Occasionally, the label equates the noogramicle, i.e. their contracted and expanded views are formally identical.
Figure 3.3 exemplifies this design. It represents the adaptive interface to a large document: the InteligˆenciaArtifical em Portugal website.8 On this design expanded noogramicles have a light gray background colour, and are on the top, whereas con- tracted noogramicles have a dark gray background colour, and are positioned below the former group. In fact all items are displayed, from top to bottom, by order of their Level of connectioness to the user request.
The noogramicle on the top is central, and represents the page for certain purposes, explained later. Actually, the values of Level shown correspond to the energy level from the spreading activation algorithm that discovers the noogramicles that are more connected to the central noogramicle (which as a reference Level of 1000). This algorithm is described fully later. On this design, the real estate on a page is divided equally between expanded and contracted noogramicles. The higher-ranking noogramicles are expanded.
Navigation, or travelling, on this design, is effected by recentring on a noogramicle. Normally, recentring is done on a non central noogramicle; but actually recentring on the central noogramicle is possible and may produce a different page, because connection levels may have changed as a consequence of travels made by other usors while this usor was (quiet) on this page.
Recentring is realised either by
• pushing the noogramicle button (the button at the start of the noogramicle labelled with the number and the contracted form of the noogramicle), or by
• selecting the noogramicle (checking the noogramicle checkbox, situated right at the start of the noogramicle) and pushing the Recenter button on the bottom of the page (to recentre on the central noogramicle only this form is available).
Actually recentring may be done on more that one noogramicle simultaneously (by selecting the respective noogramicles and pushing the Recenter button). What recen- tring does is to inkove the spreading activation algorithm with the selected noogram- icles energized, and record travel connections (explained later).
8Artificial Intelligence in Portugal, actually a branch of the website of the APPIA, Associa¸c˜ao Portuguesa Para a InteligˆenciaArtificial (Portuguese association for artificial inteligence), at www. appia.pt 3.2. MODEL DESIGN 59
Figure 3.3: Example design of our own adaptation model.
Buttons Shatter and Coalesce also work on selected noogramicles. These are authoring operations. Shatter splits up each selected noogramicle in two. Coalesce is the inverse operation.
Checkboxes Related are for relevance feedback. The usor may (or may not) check these boxes to indicate that the noogramicle is indeed related to their information need—whether the session has ended or not.
Finally, button End session serves two purposes:
• also relevance feedback, in combination with the Related checks; in particular, a session that has ended with Related checks is deemed successful
• navigation: it cleans up the interface by recentring on a neutral/start/search page.
The main adaptive input of our system consists of the choices, or travels, made by the usors. These travels are memorized in the computer as connections between the 60 CHAPTER 3. A NEW MODEL FOR ADAPTIVE HYPERTEXT respective noogramicles. The graph thus formed constitutes a model of the collective mind of usors. This modelling philosophy has been proposed and experimentally validated before in the endeavours surrounding the Principia Cybernetic Project, e.g. Heylighen 1997, Heylighen & Joslyn 2001, Bollen 2001.
This graph model of the collective mind is then automatically explored by means of spreading activation, to find the noogramicles more related to the central one (which has been chosen by the usor). Such noogramicles then form the page—the adaptive output—, in the manner already described (fig. 3.3 and surrounding text).
3.2.4 Detailed design with a network data model
To recap: a shattered document is divided into small constituends called noogramicles, which constitute the vertices of the network. The connections between the vertices are travels—representing both the original structure of the document and the recorded utilization, i.e. navigation, of the document, by its usors. So, a part of the usor model, namely, the past, known, behaviour of the usor, is contained in this database. The upper part of figure 3.4 illustrates this document model, a networked database of noogramicles (also called knowledge atoms, or simply atoms) as vertices, and travels as connections thereof.
Figure 3.5 represents the dynamic aspect of the system: a travel, effected by the user by clicking on an item of the prior page. This action represents a request for recentering on the chosen atom, which is done by serving a new page centred on that atom. Note that the travel is recorded as a new connection added to the database. The cycle then reapeats until the oracle is reached (or the session is abandoned or aborted for some reason).
3.3 Techniques and tools reused
In this section we describe spreading activation, the technique to explore network models, that we have selected for this thesis.9 Spreading activation is either
(1) the propagation of energy amongst the nodes of a network, usually formu-
9This section has its roots in Alves & Jorge 2005 “Minibrain: a generic model of spreading activation in computers, and example specialisations”, but has evolved greatly since then. 3.3. TECHNIQUES AND TOOLS REUSED 61
Figure 3.4: Relating the web page as seen and the graph model underneath.
database of atoms and travels
a page central view (not clickable)
expanded view 2
expanded view 3
contracted view 1 (label) contracted view 2 contracted view 3 contracted view 4 62 CHAPTER 3. A NEW MODEL FOR ADAPTIVE HYPERTEXT
Figure 3.5: Travelling
new travel
central view (not clickable) central view (not clickable)
expanded view 2 expanded view 2
expanded view 3 new page expanded view 3
contracted view 1 (label) contracted view 1 (label) contracted view 2 contracted view 2 click contracted view 3 contracted view 3 contracted view 4 contracted view 4 3.3. TECHNIQUES AND TOOLS REUSED 63
lated as repeated matrix by vector multiplication (Bollen 2001, Chen & Ng 2004, Crestani & Lee 2000, Dominich 2003, Huang et al. 2004, Rome 2003), or
(2) the specification of a final state of energy distribution amongst the nodes, for- mulated as a set of simultaneous equations (Anderson 1983)
Probably the two formulations are equivalent under certain conditions. However we have not studied this possibility deeply. Although Anderson 1983 seems to be the seminal item on this area, method (1) is much more featured in the literature at large. So we have settled on method (1), which we describe next.
3.3.1 A unified model of spreading activation
We have developed a generic model and software module of spreading activation, and specialisations thereof to support a number of specific models in the literature. The unification thus provided has helped us (and others, cf. Scheir & Lindstaedt 2006) understand spreading activation in general and compare specific models. We also created a new specific model, Waterline10, that reduces the number of parameters of a class of specific models.
The specific spreading activation models studied include leaky capacitor variants (Anderson 1983, Bollen 2001, Rome 2003, Huang et al. 2004), reverberative circles (Dominich 2003), Wa- terline (Alves & Jorge 2005), branch-and-bound (Chen & Ng 2004, apud Huang et al. 2004), Hopfield net (Huang et al. 2004), Contextual Network Graphs (Ceglowski et al. 2003).
3.3.2 A didactical example
We have prepared a toy data set to verify the software, which also serves to take a quick look at how spreading activation works. It is the small network of concepts illustrated in Fig. 3.6. The connections are untyped, undirected, unweighted, conveying only a general notion of association. For example, the connection PET–DOG signifies that dogs are pets, connection MARIUS–LUCKY signifies that these two entities, or NAMEs, are related (most probably Marius owns Lucky, given their species), connection LUCKY–DOG signifies that Lucky is a dog, etc.
10Waterline was initially called Watermark on the cited workshop article. 64 CHAPTER 3. A NEW MODEL FOR ADAPTIVE HYPERTEXT
Figure 3.6: Toy network, for a general understanding of spreading activation
NAME
MARIUS
LUCKY PET NINA SPECIES
DOG CAT HUMAN CLASS SNAKE
MAMMAL REPTILE
We tested variants of the LCM and Waterline models on this data. The transcript in Fig. 3.7 shows a Waterline session of three activations, or queries. Each query is launched using the following commands (technical details will be defined shortly).
Set NL Set the energy level of node N to L.
Go S Propagate S steps (or until some other stop condition occurs, as may happen with Waterline).
Each result is a list of nodes with non-null Waterline, in descending order of Waterline. The printed data represents the final state. The last query activates MARIUS, PET, and NAME as a representation of “the names of Marius’s pets”. Note how the system correctly high-ranks LUCKY and NINA. Compare with the simpler queries before. 3.3. TECHNIQUES AND TOOLS REUSED 65
Figure 3.7: Instances of spreading activation over the toy network (transcripts of sessions with the Minibrain program).
set marius 1 go 10 Level Waterline MARIUS... 0.000E+00 1.000E+00 HUMAN.... 0.000E+00 2.500E-01 LUCKY.... 0.000E+00 2.500E-01 NINA..... 0.000E+00 2.500E-01 NAME..... 0.000E+00 2.500E-01 Total.... 0.000E+00 2.000E+00 reset set marius 1 set pet 1 go 10 Level Waterline PET...... 0.000E+00 1.000E+00 MARIUS... 0.000E+00 1.000E+00 DOG...... 0.000E+00 5.000E-01 CAT...... 0.000E+00 5.000E-01 HUMAN.... 0.000E+00 2.500E-01 LUCKY.... 0.000E+00 2.500E-01 NINA..... 0.000E+00 2.500E-01 NAME..... 0.000E+00 2.500E-01 Total.... 0.000E+00 4.000E+00 reset set marius 1 set pet 1 set name 1 go 10 Level Waterline MARIUS... 0.000E+00 1.333E+00 NAME..... 0.000E+00 1.250E+00 PET...... 0.000E+00 1.000E+00 LUCKY.... 0.000E+00 5.833E-01 NINA..... 0.000E+00 5.833E-01 DOG...... 0.000E+00 5.000E-01 CAT...... 0.000E+00 5.000E-01 HUMAN.... 0.000E+00 2.500E-01 Total.... -2.220E-16 6.000E+00 66 CHAPTER 3. A NEW MODEL FOR ADAPTIVE HYPERTEXT
3.3.3 Benefits of spreading activation for information retrieval
Computer models of spreading activation have been used for cognitive studies (Anderson 1983), electric circuit simulation, and information retrieval (Bollen 2001, Crestani 1997, Crestani & Lee 2000, Dominich 2003, Rome 2003, Ceglowski et al. 2003). Naturally, here we focus on the latter field, and specifically on the respective references.
In spreading activation approaches to information retrieval the nodes represent seman- tic entities e.g. documents, terms, queries, users, sessions. The propagation represents a search for the entities that are most related—similar, relevant, associated—to the initially activated ones.
The present study helped us understand spreading activation in general, the differences and commonalities between the specific models in the literature, and we believe that both the model and the module have potential for reuse in other systems. Real or potential advantages that spreading activation offers in relation to classical content retrieval techniques include:
Natural incrementality. The network is the data model. As new data arrives the network is updated immediately. This constitutes learning. The result of spreading activation immediately reflects the new state. Therefore the model is fully incremental of its own nature. — Real advantage.11
Content type independence. “Nodes are not retrieved on the basis of specific, in- dividual node characteristics (e.g. text content of metadata), but based on their connections to other nodes. Retrieval is therefore not limited to a document’s text content and can operate on nodes containing any combination of non-text content such as audio, video, text, digitally formatted text, etc.” (Bollen 2001, p. 247) — Real advantage.
Robustness to missing information. “Activation is propagated in parallel through the network connections, which makes retrieval robust to missing information, i.e. absence of specific node information such as keyterms or labels does not nec- essarily impede retrieval.” (Bollen 2001, p. 247) — Real advantage; Huang et al. 2004
11Incrementalily is a huge issue in theories and systems of information retrieval, machine learning, and artificial intelligence in general. It is the ability to deal with new data coming in for analysis, by integration with the model already in place which was made from the old data. Naturally this requires an adaptation of the model. In recent years the respective area of study has been called data streams (Gaber et al. 2005). 3.3. TECHNIQUES AND TOOLS REUSED 67
shows an example of this, solving the scarcity problem of recommender systems.
Post-coordination. Classic information retrieval is based on precoordination: in- dexes and similarity matrices computed prior to the search (Foskett 1996). The weakness of this approach is that it is fortune-telling: the system designer must choose a finite subset of possible coordination needs. Searches falling outside this set will be inefficient or impossible. Postcoordination solves this problem in theory. In practice, the efficient implementation of postcoordination requires a departure from classical approaches. Spreading activation is one possible such alternative—in our informed intuition one of very high potential.12
Ostensive querying. “The process of retrieval is initiated by patterns of node acti- vation which allows (or even requires) querying-by-example, also referred to as ostensive querying” (Campbell 2000). (Adapted from Bollen 2001, p. 247.) — Potential advantage.
Scalability and incrementability. With classic information retrieval techniques, like Latent Semantic Indexing , the poor scalability of the singular value de- composition algorithm remains an obstacle to indexing very large collections. While techniques have been developed for making incremental updates to a scaled collection, these changes typically cannot exceed a certain threshold without trig- gering a rebuild (Ceglowski et al. 2003). With spreading activation, “for additions, the graph server simply has to parse the new documents, and add additional connections between document nodes and existing term nodes” (idem).
3.3.4 The generic model
Spreading activation is the propagation of energy amongst the nodes of a network. Our generic model, and corresponding software module, consists of means to represent the (non trivially energised part of the) network, set and inspect its state, and advance the propagation, as follows.
Propagation is based on atomic units of change, or steps. Upon each step, each source node in the network transmits a certain amount of energy to each of its targets;
12This pre-coordination problem also exists in the field of relational databases. Which indexes to create and maintain? Here the DBA (database administrator) is the fortune-teller. Advances in database research include overcoming this problem. Cf. for example the network data model developed in this very thesis. 68 CHAPTER 3. A NEW MODEL FOR ADAPTIVE HYPERTEXT these elements are defined by the user of the generic model (not the end-user of an application) as the entities Amount/2, Sources/0, Targets/1.13
Amount(X → Y ) A function returning the (possibly null) amount of energy to transmit from node X to node Y at the point of call.
Sources A representation of the (possibly ordered) set of source nodes at the point of call.
Targets(X) A representation of the (possibly ordered) set of targets of node X at the point of call.
Amount/2, Sources/0, Targets/1 are the parameters of the generic model. A con- cretisation of these elements, i.e. their realisation as arguments to the generic model, constitutes a specialisation of the generic model.
The program maintains and computes a number of energy levels, or fields, namely: Level, Gain, Loss, Waterline. Level(X) denotes the current activation level of node X. The other energy fields are auxiliary.
Upon each step, the levels of each node are updated as follows. The prime sign (0) denotes the state after the step. The absence of the sign denotes the state before the step.
Level(X)0 = Level(X) + Gain(X)0 (3.1)
X Gain(X)0 = Amount(Y → X) (3.2) Y ∈Sources(X)
X Loss(X)0 = Amount(Y → X) (3.3) Y ∈Targets(X)
Waterline(X)0 = max(Level(X)0, Waterline(X)) (3.4)
Sources/1 is an auxiliary entity derived from the model parameters Sources/0, Tar- gets/1 in the expected way:
13We use the costumary notation F/n to identify a function-like entity named F with arity n i.e. having n input parameters. 3.3. TECHNIQUES AND TOOLS REUSED 69
Sources(X) = {Y : Y ∈ Sources,X ∈ Targets(Y )} (3.5)
Equations 3.1–3.2 define the effect of a propagation step.A user, or full step, often requires pre- or post-adjustment of some levels, depending on particulars of the specific model and network state.
The module has a minimalist design, to the extent that this does not compromise good software engineering. The parameter entities refer to nodes. Reference entails existence. If a node is invoked as input that has been not referenced before, it is assumed to exist with an activation level of zero. Naturally, existence and level of a node can also be asserted explicitly.
The following entities are also provided by the module to inspect the network state.
Total(EnergyField) The total energy of the field specified (Level, Gain, Loss, Wa- terline). Alternate notation: |EnergyField|.
Range(EnergyField, LowerBound, UpperBound, Order) The possibly ordered set of nodes in the specified energy range. LowerBound, UpperBound are energy values each tagged ˙ or ˚ for closed or open respectively, i.e. as either including or excluding the value respectively. Order is either <, >, or a comma for ascend-
ing, descending, or no order respectively. Occasionally we use an Interval Field notation for Range expressions as follows:
[x < y]EnergyField = Range(EnergyField, x,˙ y,˙ <)
]x < y]EnergyField = Range(EnergyField,˚x, y,˙ <) etc.
Finally, the module provides the procedure Step to advance the propagation, i.e. to perform the state change defined in equations 3.1–3.4; and procedure Assert to energise nodes explicitly (Fig. 3.8). Usage of the module consists in a realisation of the algorithm in Fig. 3.9, in terms of the entities described. 70 CHAPTER 3. A NEW MODEL FOR ADAPTIVE HYPERTEXT
Figure 3.8: Step algorithm. Assert(x, EnergyField, z) tells the module that the node x has energy z on the specified field; in addition, Assert updates the totals and the Waterline level.
1 for each asserted node X: 1.1 Assert(X, Gain, 0) 1.2 Assert(X, Loss, 0) 2 for each X ∈ Sources: 2.1 for each Y ∈ Targets(X): 2.1.1 Assert(Y, Gain, Gain(Y ) + Amount(X → Y )) 2.1.2 Assert(X, Loss, Loss(X) + Amount(X → Y ))
3 for each X ∈ ]0, +∞[Gain: 3.1 Assert(X, Level, Level(X) + Gain(X))
Figure 3.9: Minibrain usage.
1 define Amount/2, Sources/0, Targets/1 2 assert initial state 3 while not satisfied: 3.1 pre-adjust 3.2 do one Step (Fig. 3.8) 3.3 post-adjust 4 collect results
3.3.5 About the implementation
The software module is called Minibrain. It is implemented in Ada in circa 300 lines of source code, excluding standard libraries. Internally, Minibrain uses Ada.Containers.Sorted Sets for the data structures.14 These structures use main memory, and provide state-of- the-art performance: “If N is the length of a set, then the worst-case time complexity of the Insert, Include, Replace, Delete, Exclude and Find operations that take an
14Actually, because Ada.Containers was due officially later than the time of writing, we used the well known and tried reference implementation AI302 mantained by Matthew Heaney. 3.3. TECHNIQUES AND TOOLS REUSED 71 element parameter should be O((log N)2) or better. The worst-case time complexity of the subprograms that take a cursor parameter should be O(1).” (Ada Reference Manual, A.18.9, 116/2).
All specified entities are represented rather directly. Entities that represent sets are implemented as iterators.
3.3.6 Leaky Capacitor Model (LCM)
The LCM is usually described mathematically, namely with matrix calculus, or linear algebra (Bollen 2001, Rome 2003, Huang et al. 2004).15
For a network of n nodes, a square n × n matrix M represents the propagation factors mij. Here and throughout this section, i = 1 . . . n, j = 1 . . . n.
The activation level of nodes is represented in an n-sized vector A of levels ai.A propagation step consists in updating A as follows (again, the prime indicates the state after the step).
A0 = M × A (3.6)
The matrix M requires some preparation. A propagation factor (often called connec- tion weight in the literature) combines a number of elements:
Connection weight wij The connection weight proper (this is why we don’t use the term for the combined entity). It represents the original, contingent strength, or weight, of the connection from node i to node j (if the connections are directed) or simply between i and j (if not). Let the n × n matrix W represent
these connection weights. Note wii = 0 for all i; more on this later.
Retention factor λ The fraction of energy retained by each node upon each step. Corresponds to the “speed of decay” γ in Huang et al. 2004 as follows: λ = (1−γ). The retention factor only applies to elements in the diagonal of M (see below).
Propagation efficiency α The fraction, of energy of a node, that is propagated, from that node, to each other node, on each step. Same as in Huang et al. 2004.
15Each of these sources have faults of varying severity in the respective descriptions. Hopefully on our paper Alves & Jorge 2005, integrated herein, the LCM model has been presented correctly for the first time in the literature, with the possible exception of Anderson & 1983 . 72 CHAPTER 3. A NEW MODEL FOR ADAPTIVE HYPERTEXT
Finally, M is defined as follows:
ˆ M = λ × In + α × W (3.7)
( W T if connections are directed Wˆ = (3.8) 0.5W otherwise
The diagonal of M—the connections of each node to itself—is used to represent the retention of energy along time. So for each node i, mii = λ.
If the connections are directed, and the direction of the connection is the direction of the propagation, the matrix must be transposed, in order for the multiplication to have the expected effect.
If the connections are not directed, M is symmetric. In this case, because both sides of the diagonal enter in the multiplication by A, the respective values must be halved, in order for the multiplication to produce the desired result.
The termination condition is simply a fixed number of iterations.
LCM-specialisation of Minibrain
The mathematical definition of LCM can be mapped directly to Minibrain parameters as follows:
SourcesLCM ≡ {1 . . . n} (all nodes) (3.9)
TargetsLCM(i) ≡ {1 . . . n} (ditto; argument ignored) (3.10)
AmountLCM(i → j) = mij × Level(i) (3.11)
However this is simply performing the multiplication M × A, i.e. Minibrain is being used just as a matrix calculator. This might not be the best approach for large and sparse matrices, as is usually the case with real data (for example web navigation data configures a graph with average connection degree in the region of 2 to 3). So an 3.3. TECHNIQUES AND TOOLS REUSED 73 alternate specialisation, more efficient, is one that only asserts the nodes with non-null energy, or that corresponding to non-null connections, or both.
SourcesLCMA ≡ {i : ai > 0} (nodes with non-null activation level) (3.12)
Targets (i) ≡ {j : m > 0} (nodes non-nully connected from i) (3.13) LCMM ij
The specialisation of Amount remains the same. This specialisation can still be seen as a matrix calculator, but this time optimized to sparse instances of M.
Yet another alternative is to use W , not M, in the specialisation of Targets, and effect the retention in a post-adjustment operation.
Targets (i) ≡ {j : w > 0} (nodes non-nully connected from i) (3.14) LCMW ij
Post-adjustment: Level(i) = λ × Level(i), ∀i ∈]0, +∞[Level (3.15)
This specialisation is yet more efficient because it only processes the elements in the diagonal of M that need processing (instead of the whole diagonal). This has impact for large number of nodes, irrespectively of the connection degrees.
3.3.7 Reverberative Circles (RC)
“Activation takes place according to a winner-take-all strategy. The acti- vation is initiated at the query node, and spreads over along the strongest connection thus passing on to another node, and so on. After a finite number of steps the spreading of activation reaches a node already affected (in the worst case it passes through the entire network and eventually gets back to the query node) thus giving rise to a loop called reverberative circle (as a model for short term memory). Those nodes are said to be retrieved which belong to the same reverberative circle, and they are ranked in the order in which they were traversed.”
(Dominich 2003, pages 170–171) 74 CHAPTER 3. A NEW MODEL FOR ADAPTIVE HYPERTEXT
For the RC-specialisation of Minibrain we employ the trick of equating the level of active nodes with the step count, starting at 1 for the initial query node. This allows us to use Range to both define Sources and observe the result in the prescribed order.
SourcesRC ≡ first of ] + ∞ > 0[Level (node with the highest level) (3.16)
TargetsRC(X) ≡ node to which X is most strongly connected (3.17)
AmountRC(X → Y ) = step count (3.18)
3.3.8 Waterline
We created the Waterline model to use total decay of the network as a termination condition. In the literature it has been noted that “no general methodology has been established to tune the several parameters involved in setting up an efficient spreading activation system” (Bollen 2001, p. 257)
The LCM has two parameters: M, and the number of iterations. Eliminating one would at least ameliorate the situation. The LCM with certain transmission factors, namely positive values less then 1, configures a network of decaying total energy. The Waterline is a memory associated with each node that stores the highest level attained by the node since the start of the propagation. Thus it is possible to let the propagation die its natural death, and still get results: the Waterline values.
3.4 Algorithms
Here we formally define the algorithms used in the shattered document model. The model is organized into three main operations: start page, recentring, learning. For each operation: a main, or nominal algorithm is defined, representing our proposal of shattered documents; alternative, or variant algorithms are also defined, representing alternative approaches for comparison with ours. 3.4. ALGORITHMS 75
3.4.1 Overview
In an shattered document configuration, each page is a ranked assembly of noogramicle views. The highly ranked items are expanded, the lowly ranked are contracted. Figure 3.4 illustrates this design. Normally, the noogramicles that constitute a page are the most connected to the central one, which is the single highly ranked one. They can be selected by any adequate ranking algorithm. We call this operation recentring. Our main approach, which we define here, was to apply spreading activation to the central noogramicle. Possible alternatives include Markov chains, collaborative filtering, association rules, pure chance. Actual alternatives tried, for comparison with the main approach, are Markov chains and chance.
Recentring applies normally for the ensuing pages in a session i.e. to the pages after the first, because normally the central noogramicle is identified by clicking on the respective label on the previous page. For the first, or start page, there is no chosen central noogramicle yet, so an algorithm is required to either
i select a central noogramicle without user input (and then apply recentring)
ii create an otherwise integral page
We were not able to devise a sensible way to choose a single noogramicle for method (i). We have opted for the following instance of (ii): serve the set of the items most connected globally. This page represents the page with more ability to lead to other places, globally. We call this the super page. Again we use spreading activation to construct this page. Like recentring, the start page algorithm could also be done in any of a number of ways. We tried a few, exploratorily. We chose the way defined here because we wanted to experiment with spreading activation, and exploration indicated good results. We also made a configuration with a random start page.
Finally, there is the learning operation: the network of recorded travels is updated with the user input. This is the source for adaptivity. Because each page represents a set of atoms, the user action can be translated into travels between atoms in a number of different ways. The nominal is: central to central. Other alternatives were tried and preterred. Variants retained for comparison are random learning and null learning. 76 CHAPTER 3. A NEW MODEL FOR ADAPTIVE HYPERTEXT
3.4.2 Formalization
Atoms is the full set of atoms, or noogramicles, of the entire document. In the algorithms, each atom has a memory for its current level of activation, or energy.
Active is the subset of active Atoms i.e. the atoms with a non null activation level.
Propagation is defined with the LCM model of spreading activation described in section 3.3.6; the operation “propagate” signifies one propagation step (equation 3.6). Propagation is along the recorded travels, which are directed pairs (Source Atom, Target Atom). In the algorithms, the energy is propagated along a travel either
• in the forward direction, from source to target of the travel: default case.
• backwards, from target to source of the travel: where specified.
In both algorithms, the propagation parameters α, λ (cf. session 3.3.6) are set to α = 1, λ = 1.
Not Seen Yet is the set of atoms not seen yet in the session to which the recentring operation pertains.
3.4.3 Start page algorithms
The main start page algorithm in defined in figure 3.10. All items in the entire document are energized; a small number of backward spreading activation steps take place; the most energized items represent the most connected ones. This algorithm is implemented by function Super Page in package Kasim2.Activation (body), lines
390-418 (see Appendix B), with nominal parameters Ls = 0.001,Ns = 5.
Figure 3.10: Super Page algorithm Given the required Page size:
1 activate all Atoms with the same amount Ls of energy
2 do Ns times 2.1 propagate back 3 return the Page size most active atoms in Atoms, in order of activation level 3.4. ALGORITHMS 77
3.4.3.1 Variants
The random variant for the start page is trivial: a set of randomly chosen items, randomly ranked, in a uniformely distributed way. Such algorithm is implemented by function Random Page in package Kasim2.Activation (body), lines 420–436.
3.4.4 Recentring algorithms
The main recentring algorithm is defined in figure 3.11. The selected item is activated, and energy is spread; the selected item becomes the central item of the new page; the most energized items constitute the remaining items of the new page. This algorithm is implemented by function Recentre in package Kasim2.Activation (body), lines 202–
284 (see Appendix B), with nominal parameter Lr = 1.
Loop 2 is expected to terminate under normal conditions, namely an adequate docu- ment graph and spreading activation parameters, which are provided in the implemen- tation.16 Step 3 is to ensure that Central comes up top in the resulting page. Recall that this operation is done in the context of a current session. The server is supposed to memorize the pages served (the page trail) and in particular the atoms already seen in this section, which information is required by the algorithm to implement the respective heuristic (Not Seen Yet).
Figure 3.11: Main recentring algorithm Given a Central atom and the required Page size, and with a function Selected = Active ∩ (Not Seen Yet ∪ {Central}):
1 activate Central to level Lr 2 while |Selected|h Page size 2.1 propagate 3 reactivate Central to the highest value possible 4 return the Page size most active atoms in Selected, in order of activation level
16In the implementation, also defensive programming provisions are made to ensure termination, namely the whole invocation is enveloped in a timeout construct. 78 CHAPTER 3. A NEW MODEL FOR ADAPTIVE HYPERTEXT
3.4.4.1 Variants
The random variant of recentring is trivial and equates the random variant for the start page: a set of randomly chosen items, randomly ranked, in a uniformely dis- tributed way. Such algorithm is implemented by function Random Page in package Kasim2.Activation (body), lines 420–436.
The Markov chains variants use first order Markov chains as defined on Borges & Levene 2008. We reutilize the machinery put in place for spreading activation. We observe that the first order transition probability of Markov chains is proportional to the travel count, so spreading once from the anchor will energize the candidate items adequately. A single propagation step is taken. The page size might be incomplete. The pure Markov chains algorithm is defined in figure 3.12. The Markov chains algorithm incorporating the same Not Seen Yet heuristic as the nominal algorithm (figure 3.11), is defined in figure 3.13. The algorithms are implemented by function Recentre in package Kasim2.Markov (body lines 74–124), and function Recentre in package Kasim2.Markov With Heuristics (body lines 74–148), respectively (see Appendix B).
Figure 3.12: Pure Markov chains recentring algorithm Given a Central atom and the maximum Page size:
1 activate Central to level Lr 2 propagate 3 reactivate Central to the highest value possible 4 return the Page size (or less) most Active atoms, in order of activation level
Figure 3.13: Heuristical Markov chains recentring algorithm Given a Central atom and the maximum Page size, and with a function Selected = Active ∩ (Not Seen Yet ∪ {Central}):
1 activate Central to level Lr 2 propagate 3 reactivate Central to the highest value possible 4 return the Page size (or less) most active atoms in Selected, in order of activation level 3.4. ALGORITHMS 79
3.4.5 Learning algorithms
In our system, learning consists in recording the travels made by the usors. Simple as this concept seems, it poses a problem in a shattered document environment: since the pages are made of atoms, which connections between which atoms should represent a travel between pages? Given that we already have the notion of central atom that is a kind of representative of the page, we intuited that travelling should be recorded as a connection between the central atoms travelled from and to, i.e. as a connection from the central atom of the source page (the page where the user has clicked) to the central atom of the target page (the atom clicked upon). This central to central strategy is our main learning strategy, which effectiveness we have verified against the obvious alternatives. All are defined as follows, and coded as enumeration Update Travels T on package Kasim2 (spec), lines 184–192. The respective logic is implemented in the body of procedure Simulate Session on lines 149–221 of package Kasim2 (body).
Central To Central Increment the travel count from the current central atom to the current choice = next central atom. (Default behaviour.)
Page To Central Increment the travel count from the current entire page (all atoms except choice) to the current choice = next central atom.
All Previous Central To Central Increment the travel count from all previous and current central items to the current choice = next central atom.
All Previous Page To Central Increment the travel count from all previous and current pages, from all items of each page (except the current choice), to the current choice = next central atom.
Randomly Increment the travel count from a randomly chosen atom to a randomly chosen atom in a uniformely distributed way. This variant serves to test the respective random configuration using the same machinery put in place for all configurations.
Central To Page Increment the travel count from the central atom to all items of the next page.
Page To Page Increment travel count from each page item (except choice) to each item of next page. A lot of connections viz. (N-1)*N; normally N=10; 90 connections. 80 CHAPTER 3. A NEW MODEL FOR ADAPTIVE HYPERTEXT
False No increment. No learning. This variant serves to test non-adaptive configura- tions, notably the original structure of the document, using the same machinery put in place for all configurations.
3.5 Summary
In this chapter we have defined the principal algorithms of the shattered document approach—Start Page, Recentre, Update Travels, and the algorithms of spreading activation used to implement them. Next we will define the experimental setup including the simulator. Chapter 4
Experimental methodology
When you can measure what you are speaking about, and express it in numbers, you know something about it. Lord Kelvin
We have defined the experimental methodology described in this chapter, with the objective of evaluating our approach. We have proceeded via simulation i.e. we have designed a user simulator, based on existing findings about hypertext user be- haviour (Ahn et al. 2005, Bollen 2001, Sosnovsky & Brusilovsky 2005, Olson & Chi 2003, West et al. 2009).
The advantages of a simulation methodology over a live user study include the ability to experiment with many different configurations or parameters, and absolute control over all experimental conditions—namely assuring that the conditions that do not pertain to the variants under comparison are exactly the same. The disadvantages of a simulation methodology over a live user study include the fact that the simulation is a simplification of the reality—with the risk of being an over-simplification—and the reliability of the results depends on the proper calibration of the simulator—which can only be approximated and based on theory, in the absence of live data.
Live user experiments are essential for hypertext usabilitity studies. We have used simulation based on existing results from live studies. We wanted indeed to exper- iment with many different configurations of hypertext, and parameters of spreading activation. We hoped to conduct a proper user study too. The construction of the
81 82 CHAPTER 4. EXPERIMENTAL METHODOLOGY simulator and of the adaptation software, and the many experiments done therewith, has turned out to attain an ammount of work such as to preclude the realisation of the user study within the limits of the thesis. It is now hoped for future work (see chapter 6 Conclusions).
4.1 Simulation
To validate our hypotheses we have conducted a number of experiments via simulation. For that, we have designed and implemented a simulator, along with some other necessary programs. The design of the simulator is an interesting contribution on its own. We intend to emulate a set of usors utilizing a large document via an hypertext interface. The whole system is comprised of a large document, the usors of that document, and usor sessions whereby the usors navigate the document to pursue an information need. The concept of session is the crux of the system, as it connects the two models of document and usor. This is done via a certain model of usor behaviour, or navigation, and a transposition of the vague concept of information need to the tangible construct of an oracle (the noogramicle that holds the sought information).
We have designed the simulator of usor sessions—a type of construct sometimes called user model in the literature—, integrating theoretical and experimental re- sults from Ahn et al. 2005, Bollen 2001, Sosnovsky & Brusilovsky 2005, Olson & Chi 2003, West et al. 2009 and others. The main design premise might be called the Smart User Assumption, whereby the user chooses the right link, asymptotically. This assumption is related to Clear Labels (section 3.1.2). Support for this assumption is as follows. For the case when the oracle is only one click away, the label clearly identifies the oracle, and therefore the user selects the item easily.
When the oracle is further away than one click, the intelligence or intuition of the user takes place to select the item most likely to lead to the goal.1 Bollen 2001 has demonstrated that users successfully apply a hill climbing strategy to navigate in a web of words. Olson & Chi 2003 provides additional support for the fact that
1In reality, this is not always true. Usors make errors, or the labels are not always clear enough, or both. A possible refinement to our model could be to associate a probability of the user to choose the best link to follow. Additionally, the simulator would require the possibility for the user to backtrack. However, we are conviced these hypothetical faults or approximations of the simulator do not affect the comparative evaluation. Recall that exactly the same simulator is used for all configurations of adaptive hypertext under experiment. 4.1. SIMULATION 83 users are good sniffers. West et al. 2009 shows that hypertext users “leverage semantic associations based on background knowledge of many common sense facts, and select links according to this knowledge”. And let us recall from chapter 2 how searchers use “a combination of ideas such as what the target information might look like, where it might be found, or how one might go about tracking it down” (Campbell 2000). All these results have been obtained from controlled experiments with real users. They have provided inspiration and support for the simulator.
Another fundamental trait of the simulator is the concretization of the information need of the usor in a session as an oracle. An oracle is a small set of noogramicles, normally just one, holding the information sought by the usor in a session. The simulator supports a statistical model of oracles, based on a power law distribuition of their popularity: a few oracles are very popular, many oracles have low popularity. Actually in the final experiments we have used a uniform distribution of a subset of atoms as the oracle model, as an approximation of the power law distribution of all; exploratory experimentation with both types of distribution has indicated that the former is indeed a good enough approximation of the latter in our system i.e. they showed similar results.
4.1.1 Formalization
We reuse formal entities introduced in the formalization of the main algorithms of shattered documents (section 3.4.2).
The user is simulated by the function Choose (figure 4.1) that, given a page in a session, identifies the item on that page that the user will click on. Because we are using simulation, we know the oracle of each session. From the Smart User Assumption, the user clicks on the item most likely to lead to the oracle. We interpret this likelihood as connectedness in the travel graph, and again use spreading activation to explore this connectedness: the oracle is energized and its energy is back propagated until (at least) one link of the page being seen is activated.
The Choose algorithm in formally defined in figure 4.1. This algorithm is implemented in function Choose of package Kasim2.Activation (body), lines 32–110 (see Appendix
B), with nominal parameters α = 1, λ = 1, LZ = 1.
The loop 2.2 is theoretically guaranteed to terminate, under normal conditions, namely a fully connected graph and no decay (λ ≥ 1): eventually all atoms will receive energy, 84 CHAPTER 4. EXPERIMENTAL METHODOLOGY negating condition 2.2.1 trivially. Nevertheless the costumary defensive programming provisions were included in the implementation.
Figure 4.1: Choose algorithm Given a page of Page Atoms, of which some are Contracted, in a session with oracle Z, and with the function Selected = Active ∩ Page Atoms:
1 if Z ∈ Contracted then 1.1 return Z 2 otherwise
2.1 activate Z to energy level LZ 2.2 while Selected = ∅: 2.2.1 propagate back 2.3 return the most active Selected atom
4.1.1.1 Variants
A random variant of Choose was prepared for verification purposes. The algorithm is trivial: pick a random page item (non-central), in a uniformely distributed way. Such algorithm is implemented by function Random Click in package Kasim2.Comparate.Configurations (body), lines 12–15.
4.2 Experiments
Firstly we have conducted verificatory and exploratory small scale experiments, with artificial documents. Secondly we have realised real life size experiments with a real document, the Ada Reference Manual, or ARM.
The first cycle helped us understand the parameters involved, and provided indicative measures of performance of our approach. These indicative results were encouraging. Figure shows the evolution of session size for ten oracles in an artificial document of 60 atoms. This first cycle of experimentation is fully described on appendix C.2. 4.3. PARAMETER SETTINGS 85
Figure 4.2: Evolution of session size in an exploratory experiment
These preliminary results are a good illustration of the ability of the method we propose to reduce the user effort, here measured as session size. However, to test our hypotheses, we will use a more robust setup, which we describe henceforth on this chapter. We intend to make a comparative evaluation of several approaches to adaptive hypertext. Namely, we intend to compare:
• an adaptive system against one without adaptation
• the shattered documents approach against the (unchanged) traditional structure for documents
• the spreading activation method against another, more standard, method (Markov chains)
4.3 Parameter settings
The experiments consist in the generation of sessions, using the simulator, for different settings, or configurations. The system comprises two models (fig. 4.3). 86 CHAPTER 4. EXPERIMENTAL METHODOLOGY
Figure 4.3: Top level models and respective components of the experimental setup. The subsidiary nodes represent the main components of each model.