Generalizable Semantic Relations for Knowledge Representation

Transcription

1 Generalizable Semantic Relations for Knowledge Representation CTF 07 Conference Düsseldorf, August 2007 Katrin Weller, Isabella Peters, Wolfgang G. Stock Institute for Language and Information, Dep. of Information Science, Heinrich-Heine-University Düsseldorf

2 Outline 1. Knowledge Representation in Information Science 2. Classical Relations in Knowledge Representation Methods 3. Syntagmatic Relations in Folksonomies 4. Construction of Relations in Ontologies CTF 07 - Generalizable Relations in Knowledge Representation 2

3 Practical Background: Content Indexing and Knowledge Representation Knowledge representation from information science s perspective: General aim: Retrieval of (relevant) information and embedding in current work flows. Solution: Indexing Documents are provided with descriptive metadata. Methods of content indexing and information retrieval act together. Knowledge representations are constructed for practical aims CTF 07 - Generalizable Relations in Knowledge Representation 3

4 Practical Background: Content Indexing and Knowledge Representation Building knowledge representation models Development of classification schemes, thesauri, rules for abstracting, etc. Information retrieval Search with controlled vocabularies, relevance judgments via given abstracts or metadata. Content Indexing of Documents E.g. classification, annotation, abstracting CTF 07 - Generalizable Relations in Knowledge Representation 4

5 Knowledge Representation Methods in Information Science Methods for knowledge representation Classifications Thesauri Controlled keyword indexing (Schlagwortmethode) Methods differ in complexity (e.g. how concepts can be interrelated). Some methods have been standardized, for some there are national or international norms (e.g. DIN for classification systems) CTF 07 - Generalizable Relations in Knowledge Representation 5

6 Classification System: Example IPC, International Patent Classification CTF 07 - Generalizable Relations in Knowledge Representation 6

7 Retrieval Without Controlled Vocabulary Example: Google Example: Flickr Examples from and CTF 07 - Generalizable Relations in Knowledge Representation 7

8 Retrieval Without Controlled Vocabulary Basic aim of information retrieval: find what I mean, not what I say. Example I: Searching for photos of Düsseldorf, with search term Düsseldorf. Not retrieved are: Examples from CTF 07 - Generalizable Relations in Knowledge Representation 8

9 Retrieval Without Controlled Vocabulary Example II: Search for bank retrieves: German Bank = bench District Bank in London Examples from CTF 07 - Generalizable Relations in Knowledge Representation 9

10 Practical Use of Controlled Vocabularies The terminology of a domain of knowledge is represented as a structured model of concepts and relations. Bundling of synonyms, separation and explanation of homonyms. Reduction of varieties. A generally valid and consistent way of indexing is enabled. Unification of users and indexers vocabulary. A search query can be refined and expanded with additional query terms (query expansion or query reformulation). Search results can be clustered according to contexts CTF 07 - Generalizable Relations in Knowledge Representation 10

11 Relations in Knowledge Representation Paradigmatic Relations: fixed, rigidly coupled concept relations within controlled vocabularies. Example: vehicle and bicycle as hierarchy within a classification scheme. Syntagmatic Relations originate merely in the actual cooccurrence of terms within a certain setting. Generalizable Relations are paradigmatic relations that can be meaningfully used in all (or in most of all) general or domain-specific knowledge representation and organization models CTF 07 - Generalizable Relations in Knowledge Representation 11

12 Research on Relations There are detailed and highly-professional theoretical reflections on relationships from fields such as linguistics, philosophy, information science, knowledge engineering, artificial intelligence studies and related disciplines. Also to be considered: properties, attributes-value pairs, slots and fillers... New Approach: Examination of existing knowledge representation approaches. Application of detected relations to practical tasks CTF 07 - Generalizable Relations in Knowledge Representation 12

13 Classical Relations Used in Knowledge Representation Systems In contrast to rich theoretical research, only a small number of paradigmatic semantic relations is currently differentiated and practically applied in classical methods of knowledge representation: Classical Types of Relations Relation of equivalence Hierarchical Relations Hyponymy Meronymy Associative relations (all other relations) CTF 07 - Generalizable Relations in Knowledge Representation 13

14 Classical Relations Used in Knowledge Representation Systems Relation of equivalence Synonymy (Relation between different names for the same concept) Quasi-Synonyms (Relation between concepts with similar meaning) Genidentity (Relation between concepts, whose meaning has changed slightly in the course of time (Leningrad - St. Petersburg) Using this relations helps to unify a controlled vocabulary and to increase recall in information retrieval CTF 07 - Generalizable Relations in Knowledge Representation 14

15 Classical Relations Used in Knowledge Representation Systems Hierarchical Relations Relations between (upper) concepts and their sub-concepts. Basis framework for most concept models dominate current systems. Hyponymy Logical view. Specification of features; is a kind of, is_a. kind-of-relation, taxonomic relation, taxonomy Examples: mammal cat; vehicle car. Meronymy Objective view; is a part of. Part-whole-relation, mereology, part-of relation, partonomy (meronym holonym). Examples: car motor; England London CTF 07 - Generalizable Relations in Knowledge Representation 15

16 Classical Relations Used in Knowledge Representation Systems Associative Relations They are unspecified connections of concepts that can have any kind of relation except synonyms and hierarchical relations. Proposals for specification: contrasts (antonymy) cause effect (causality) producer product (genetic relation) material product predecessor successor (succession) sender recipient (transmission) etc CTF 07 - Generalizable Relations in Knowledge Representation 16

17 Classical Relations Used in Knowledge Representation Systems Complexity in structure Thesaurus Classification Hyponymy, meronymy, equivalents & associations Hierarchy & equivalents Controlled Keywords Equivalents & associations Extend of captured knowledge domain CTF 07 - Generalizable Relations in Knowledge Representation 17

18 Use of Relations: Query-Expansion Search queries can be expanded by adding concepts related to the initial search terms to improve search results. Example: Search for a stud farm in the German region Rhein-Erft-Kreis. region Rhein-Erft-Kreis cities Bergheim Kerpen districts Quadrat- Ichendorf Niederaußem CTF 07 - Generalizable Relations in Knowledge Representation 18

19 Reconsidering Classical Relations Caution! Expansions may include only one path, if the relations are not transitive. Transitivity Example: Tottenham is_part_of London; London is_part_of UK. And also: Tottenham is_part_of UK. Classical counter-example: Nose is_part_of Professor; Professor is_part_of University. But not: Nose is_part_of University. Possibly solution: specification of meronymy CTF 07 - Generalizable Relations in Knowledge Representation 19

20 Decomposition of Classical Relations Example: Meronymy Meronymy (part-whole) Decomposition of structures Geograph. subunit geograph. unit Structure-independent composition Piece - Whole Element - Collection Phase - Activity Unit - Organization Segment - Event Component-Complex Part - Object Portion -Compound CTF 07 - Generalizable Relations in Knowledge Representation 20

21 New Methods for Knowledge Representation Current Trends Web 2.0 Knowledge Representation with Folksonomies. Users take over the indexing of web content. Semantic Web Knowledge Representation with Ontologies. Higher complexity, formal structuring. Examples from and CTF 07 - Generalizable Relations in Knowledge Representation 21

22 New Methods for Knowledge Representation Complexity in structure Trend: enhancement Ontology Thesaurus Hyponymy, meronymy, equivalents & specified associations Classification Hyponymy, meronymy, equivalents & associations Hierarchy & equivalents Controlled Keywords Folksonomy Equivalents & associations (no paradigmatic relations) Extend of captured knowledge domain CTF 07 - Generalizable Relations in Knowledge Representation 22

23 Folksonomies Content indexing 2.0 Users have begun to organize web content (e.g. photos, bookmarks). Tags are added to documents to describe their content. The process is called (social) tagging and resembles uncontrolled keyword indexing. The whole collection of tags within one platform is called folksonomy. Tags can then be used for searching CTF 07 - Generalizable Relations in Knowledge Representation 23

24 Relations in Folksonomies? No controlled vocabulary, no explicit relations. Like in plain texts, relations in folksonomies are purely syntagmatic. But: Users Terminology is captured and can provide insights to their actual language use for indexing/searching. Relations may be hidden in the combined assignment of several tags. Further research is needed CTF 07 - Generalizable Relations in Knowledge Representation 24

25 Example I (geographic) Hierarchy Synonyms Is_A CTF 07 - Generalizable Relations in Knowledge Photos Representation and tags from

26 Example II Location - (geographic) hierarchy with region Equivalent: football - soccer Name of Stadium changes over time Related colours Rival Photos and tags from CTF 07 - Generalizable Relations in Knowledge Representation 26

27 Example III del.icio.us: Bookmarks can be tagged by different users: tag clouds. Examples from: CTF 07 - Generalizable Relations in Knowledge Representation 27

28 Evaluation of Social Tags Systems, where documents are tagged by several users, tend to cover synonyms and spelling variants. We can also find kinds of hierarchies. Often more than one tag is assigned to each document, so that interrelations may be exploited. Differences can be noted between tags referring to pictures and documents/bookmarks. Tags also include many personal associations, most of them cannot be used for our purpose (e. g. me, to do, my dog, last year, fantastic, readthis) CTF 07 - Generalizable Relations in Knowledge Representation 28

29 Ontologies Ontologies as new knowledge representation systems consist of concepts, instances, relations that connect concepts ans specify properties of instances. Ontologies can be represented with formal ontology languages such as OWL. Ontologies shall enable the semantic annotation and integration of information in a Semantic Web. Currently there are no rules or guidelines regarding the use of relations in ontology engineering CTF 07 - Generalizable Relations in Knowledge Representation 29

30 Elements of Ontologies Classes: hierarchical editor, rules for class membership can be added. Instances (individuals) are assigned to classes Excerpt CTF from07 Generations - Generalizable Ontology: Relations in Knowledge Representation 30

31 Elements of Ontologies Additional relations can be chosen to model the domain. Object and datatype relations. Relations may be specified: Domain and range Functional? Symmetric? Transitive? Inverse counterpart Cardinalities CTF 07 - Generalizable Relations in Knowledge Representation 31

32 Example Generations Ontology takenctf from: Generalizable Relations in Knowledge displayed Representation with Protege OntoViz Tab. 32

33 Relations in Ontologies Most dominant: hierarchical is_a, instance_of; also part_of. But: Self-defined relations should be used to capture the domain appropriately. Many relations are domain-specific. Question: Are there new relations that apply to most or many domains of interest (and are these used in current ontologies)? CTF 07 - Generalizable Relations in Knowledge Representation 33

34 Example: Domain Specific Relations DOLCE : a Descriptive Ontology for Linguistic and Cognitive Engineering. Section Social Units. Example: DOLCE (Section Social Units, CTF 07 - Generalizable Relations in Knowledge Representation 34

35 Example II UMLS Unified Medical Language System Set of non-hierarchical relationships, grouped into five major categories: physically related to spatially related to temporally related to functionally related to conceptually related to See: and CTF 07 - Generalizable Relations in Knowledge Representation 35

36 Suggestions for Generalizable Relations Properties for General Use? Location: geographical information (including hierarchies), locations in comparison to other objects (inclusion, adjacency) Perception: color, shape, taste, sound, haptics, surface, texture, smell Measures: Length, weight, height, density, temperature, time (duration), material?... Functions, actions, task, behavior, processes, events, motions: e.g. develops_from, used_for, used_by, produced_by, transformations, motions CTF 07 - Generalizable Relations in Knowledge Representation 36

37 Conclusions Current methods of knowledge representation for information organization and retrieval use only limited types of relations. Differentiated relations that can be explicitly formalized in ontologies or may be inherently hidden in folksonomies. More relations will be collected, clustered and characterized. Newly identified general-use relations will than have to be evaluated for their use in practical applications CTF 07 - Generalizable Relations in Knowledge Representation 37

38 References I Bean C.A., Green R. (ed.): Relationships in the Organization of Knowledge. Dordrecht: Kluver, Bertram, J.: Einführung in die inhaltliche Erschließung. Grundlagen, Methoden, Instrumente. Würzburg: Ergon, Cruse, D.A.: Hyponymy and its varieties. In: Green, R., Bean, C.A., & Myaeng, S.H., ed.: The Semantics of Relationships. Dordrecht: Kluwer, 3-21, Gerstl, P., Pribbenow, S.: A conceptual theory of part-whole relations and its applications. In: Data & Knowledge Engineering, 20: , Green, R., Bean, C.A., & Myaeng, S.H., ed.: The Semantics of Relationships. Dordrecht: Kluwer, Gordon-Murnane, L.: Social bookmarking, folksonomies, and Web 2.0 tools. Searcher - The Magazine for Database Professionals, 14(6), 2006, Guy, M., Tonkin, E.: Folksonomies: Tidying up tags? D-Lib Magazine, 12(1), Hovy, E.: Comparing Sets of Semantic Relations in Ontologies. In: Green, R., Bean, C.A., & Myaeng, S.H., ed.: The Semantics of Relationships. Dordrecht: Kluwer, , CTF 07 - Generalizable Relations in Knowledge Representation 38

39 References II Lancaster, F.: Indexing and Abstracting in Theory and Practice. 2nd ed, London: Library Association Publishing, Löbner, S.: Understanding Semantics. London: Edward Arnold Publishers, Khoo, C.S.G. & Na, J.C.: Semantic relations in information science. In: Annual Review of Information Science and Technology, 40, 2006, Mathes, A. (2004). Folksonomies Cooperative Classification and Communication Through Shared Metadata. Retrieved August 15, 2006 from Pribbenow, S.: From classical mereology to complex part-whole-relations. In: Green, R., Bean, C.A., & Myaeng, S.H., ed.: The Semantics of Relationships. Dordrecht: Kluwer, 35-50, Stock W.G.: Information Retrieval. Suchen und Finden von Informationen. München; Wien: Oldenbourg Wissenschaftsverlag, W3C Recommendation: OWL Web Ontology Language Overview, 2004, (retrieved March 30, 2007). Winston, M.E., Chaffin, R., & Herrmann, D.: A taxonomy of part-wholerelations. In: Cognitive Science, 11: , CTF 07 - Generalizable Relations in Knowledge Representation 39