0 Semantic Web - (2019) 0 IOS Press 1 1 2 2 3 3 4 Are We Better Off With Just One Ontology 4 5 5 6 on the Web? 6 7 7 8 Armin Haller a,* Axel Polleres b 8 9 a Research School of Management & Research School of Computer Science, Australian National University, 9 10 Canberra, Australia 10 11 E-mail: [email protected] 11 12 b Institute for Information Business, Vienna University of Economics and Business, Vienna, Austria 12 13 E-mail: [email protected] 13 14 14 15 15 Editors: Pascal Hitzler, Wright State University, USA; Krzysztof Janowicz, University of California, Santa Barbara, USA 16 16 17 17 18 18 19 19 20 20 Abstract. Ontologies have been used on the Web to enable semantic interoperability between parties that publish information 21 21 independently of each other. They have also played an important role in the emergence of Linked Data. However, many ontologies 22 on the Web do not see much use beyond their initial deployment and purpose in one dataset and therefore should rather be called 22 23 what they are – (local) schemas, which per se do not provide any interoperable semantics. Only few ontologies are truly used as 23 24 a shared conceptualization between different parties, mostly in controlled environments such as the BioPortal. In this paper, we 24 25 discuss open challenges relating to true re-use of ontologies on the Web and raise the question: “are we better off with just one 25 26 ontology on the Web?” 26 27 27 Keywords: Ontology, Knowledge Representation 28 28 29 29 30 30 31 31 32 1. Introduction – Confounding conflicts refer to those arising from 32 33 1 the confounding of concepts which are in fact dis- 33 Back in 1993, Gruber introduced “ontologies” as tinct. An example is the maximum temperature on 34 explicit specification of a conceptualization 34 an “ ” con- a given day. Due to different time-periods (e.g., 35 sisting of a “set of objects, and the describable re- 35 36 calendar day vs. a 24 hour time-period) and dif- 36 lationships among them” represented in a declarative ferent methods of averaging (e.g., over a minute 37 37 formalism [23]. Uschold and Grüninger [65] argued vs. over an hour) the actual values, even when 38 38 later that semantic interoperability between parties that recorded by the same sensor, will often differ 39 39 exchange data is a key application of ontologies. when published by different parties. 40 40 The use of ontologies as an approach to overcome – Naming conflicts occur when naming schemes 41 41 the problem of semantic heterogeneity on the World of information differ significantly, for example 42 42 Wide Web has since been well established. Semantic synonyms and homonyms among attribute values. 43 43 heterogeneity occurs whenever two contexts do not use For example, the entities Product and Item are of- 44 44 the same interpretation of information. According to ten found to be synonyms in commerce applica- 45 45 Goh [21] three causes for such semantic heterogeneity tions. 46 46 can be identified. – Scaling and units conflicts refer to the adoption 47 47 of different units of measure or scales, e.g., impe- 48 48 *Corresponding author. E-mail: [email protected]. rial gallon vs US gallon vs litre. 49 49 1The plural use of the term “ontology” in computer science quite 50 likely still raises eyebrows for anyone with a background in ontology Many ontology-based approaches that address these 50 51 in philosophy. causes of semantic heterogeneity have been proposed 51 1570-0844/19/$35.00 c 2019 – IOS Press and the authors. All rights reserved A. Haller and A. Polleres / Are We Better Off With Just One Ontology on the Web? 1 1 since [47, 71]. The idea is that a shared ontology which initions of terms. Examples of domain ontolo- 1 2 carries a formal semantics, acts as a gold standard for gies are the Gene Ontology [3], the Disease On- 2 3 the definition of information in different contexts and tology [57], ChEBI [15], the Building Topology 3 4 applications. Many kinds of ontologies have been pro- Ontology (BTO) [55] or VSSo [38], the Vehi- 4 5 posed that can be classified on a spectrum from very cle Signal and Attribute Ontology. The latter is 5 6 lightweight ones that may consist of terms only, with a recently developed car signal ontology that de- 6 7 little or no specification of the meaning of the term, to rives from the automotive standard VSS, and that 7 8 rigorously formalized logical theories [66]. In this pa- builds upon a mid-level ontology pattern, i.e., 8 9 per we focus on the latter, i.e., formal ontologies ex- from SSN/SOSA, for representing observations 9 10 pressed in RDFS/OWL. and actuations. 10 11 The ontology engineering community has proposed 4. Use case ontologies include a set of detailed 11 12 ontologies with different levels of abstractions to ease classes and relations highly dependent on the use 12 13 reuse and to also layer ontologies upon each other. case. For example, in a smart home environment 13 14 Although no agreed upon ontology hierarchy exists, for an apartment building, a use case ontology 14 15 adapting the ontology classification of Guarino [25], may extend terms in a domain ontology to be 15 16 we can largely distinguish four different levels of ab- able to use those terms for a number of similar 16 17 straction in ontology design as shown in Figure1. units in an apartment complex. 17 18 18 19 19 20 Highest 2. Challenges in Reusing Ontologies 20 e.g. CyC, SUMO, Upper General Most 21 DOLCE, BFO, CYC 21 Ontologies 22 While upper ontologies have experienced strong re- 22 e.g. PROV-O, FOAF, 23 ORG, SOSA/SSN search interest in the early 2000’s, their use on the 23 Mid-Level Ontologies 24 Web has largely been confined to the biomedical do- 24 Reusability e.g. GO, ChEBI, 25 DO, BTO, VSSo Abstraction of Level main where the community, through the OBO foundry, 25 Domain Ontologies 26 maintained and mandated the use of the BFO upper on- 26 27 tology. In fact, in an analysis of links [29] in the LOD 27 Use-Case Ontologies Most SpecificMost 28 Lowest Cloud [1] we have discovered that not a single dataset 28 29 in a corpus of 430 Linked Open Datasets that were in- 29 Fig. 1. Levels of Abstraction in Ontology Design 30 vestigated for this study reuses DOLCE or SUMO, the 30 31 other two main open-source upper ontologies. 31 32 1. Upper ontologies that define very general terms This lack of adoption of upper ontologies outside 32 33 that are common across all knowledge domains, the biomedical domain can mostly be attributed to 33 34 examples of which are CYC [40], SUMO [46], the complexity and rigidity of these ontologies and 34 35 DOLCE [19] and BFO [60]. the often unintended inferences that would result from 35 36 2. Mid-level ontologies (sometimes also called top importing the upper ontology in a mid-level or do- 36 37 domain ontologies or global domain area ontolo- main ontology. Examples of such unintended infer- 37 38 gies) act as a bridge between the abstract con- ences are global domain and range restrictions de- 38 39 tent of an upper ontology and the richer detail of fined in an upper ontology (e.g., DOLCE+DnS Ultra- 39 40 various domain ontologies. Space and time are lite (DUL) uses global property restrictions) that may 40 41 two modelling aspects shared between any do- lead to inferences in the importing domain ontology 41 42 main, and ontologies such as the OWL Time On- that are inconsistent in its domain of discourse. An- 42 43 tology [9] and Geonames are widely used across other example is the disjointness of a set of classes 43 44 domains. Other examples of mid-level ontolo- defined in an upper ontology that results in an unin- 44 45 gies are PROV-O [39], FOAF [6], ORG [56] and tended restriction on the use of the domain class that 45 46 SOSA/SSN [30] that define concepts generally is a subclass of such an upper level class. For exam- 46 47 enough so that their semantics can be further nar- ple, in the old SSN, the Sensor class was defined 47 48 rowed by a domain ontology. as a subclass of a DUL PhysicalObject. How- 48 49 3. Domain ontologies define concepts and relations ever, users of the SSN ontology who wanted to use 49 50 that belong to a specific domain. Each domain the Sensor class for computational methods, could 50 51 ontology typically models domain-specific def- not, because a dul:PhysicalObject is disjoint 51 2 A. Haller and A. Polleres / Are We Better Off With Just One Ontology on the Web? 1 with a dul:SocialObject (which most certainly In the following we identify a set of challenges 1 2 would include a computational algorithm). For this and that we have repeatedly encountered in ontology en- 2 3 other reasons [30], in the redesign of the SSN ontol- gineering consultancies with Government and indus- 3 4 ogy, the working group decided to remove the depen- try clients. These include some of the ontology evalua- 4 5 dency of the SSN ontology on the DOLCE Ultralite tion criteria above (some of which, e.g., clarity, consis- 5 6 ontology and make its alignment optional, i.e., provide tency, correctness, conciseness, are combined together 6 7 it in a separate ontology file that is not imported [30] into one category, ‘quality’), but also include other 7 8 (while at the same time relax its semantics by using challenges that are specific to the reuse of distributed 8 9 higher level ontology classes from DUL).