Matching Natural Language Data with Ontologies

Abstract

Ontologies and natural languages are complementary. Whereas ontologies are used to model knowledge formally, natural language is primarily used by users to communicate with ontology based systems. In order to transform information or queries in natural language into valid ontological expressions, the meaning of natural language entities have to be matched with the given ontologies. In contrast to pure ontology matching, the matching with natural language data poses some problems linked to their ambiguities (synonymy, homonymy/polysemy, redundancy, to name but a few).

-

Apart from the ontologies, the matching requires a complete lexicon of the language used to label or describe the ontological classes and properties (= entities). Our lexicon is lso linked to a semantic thesaurus. The ontologies, on the other hand, usually have non-ambiguous entity labels (like http: //www.acemedia.org/ontos/tennis#Player1) or a comment, explaining the entity. This is especially necessary if the entity labels are not self-explanatory like 1 We shorten name spaces like http://www.acemedia.org/ontos/tennis# to “tennis:” etc. tennis:C12 (a fortunately rare case). Further, the semantic thesaurus contains a thematic hierarchy of all semantic concepts to help disambiguation. These are grouped into 880 themes which in turn are organized in 80 domains. Domains are divided into about 10 macro-domains. The matching itself comprises several lexicon semantic thesaurus (e) adding ontologies to  semantic thesaurus domain ontologies (a) extract and correct  class and property labels

NLP (b) Semantic analysis of  classes (taking into  account their taxonomic 

context) (c) Semantic analysis of  properties (taking into  account their range and  domain) (f) Creation of semantic  steps (cf. fig. 1). Apart from a (more or less ccluosmteprlienxg  crulalessse fsor  manual) preparation in order to correct possi(g) Creation of RDFS  ble labeling errors in the ontologies, the other rules (predicate­class  steps do not need any intervention: (a) extracttransformation)

ing the “ontological context” of entities and as(d) Daentedc htiynpgo snyynmosnyms  signing eventual reformulations of entity labels; Fig. 1. linguistic-ontological matching (b) natural language processing passes: detecting meanings for classes using their ontological context (direct sub-classes); (c) and for properties using their ontological context (domain and range classes); (d) determining the application depending synonyms and co-hyponyms; (e) adding the ontological hierarchy to the semantic taxonomy; (f) creating semantic transformation rules for “complex class labels”2; (g) creating transformation rules for the creation of ontological representation (from semantic graphs. Synonyms (defined in our multilingual thesaurus, [ 9 ]) are all matched onto the same ontological class (e.g. “river”, “stream”, “creek” etc. → holidays:River). If a class has no sub-classes, we also match the co-hyponyms of the label to the class (e.g. in our case “car”, “bus”, “truck”, “motorbike” . . . → general:Vehicle. The resulting linguistic data is successfully used the aceMedia prototype, similarly produced data is used in an industrial application to create and access ontological based information from/via natural language.

New perspectives are offered by structured semantic data which is getting more and more available. Databases like Wikipedia (especially the categorization schema used within) or RDF or ontology based information systems like DBpedia or freebase3 (both initialized by Wikipedia contents) will help to improve the linking of natural languages and formally modeled ontologies. 2 Labels which use multi-word expressions like tennis:ExhibitionMatch instead of simple words. 3 http:// dbpedia.org/ , http:// freebase.com/

References

1. Heinecke , J.: Ge´ne´ration automatique des repre´sentations ontologiques . In: TALN . ( 2006 ) 502 - 511

2. Dasiopoulou , S. , Heinecke , J. , Saathoff , C. , Strintzis , M.G. : Multimedia reasoning with natural language support . In: IEEE-ICSC . ( 2007 ) 413 - 420

3. Heinecke , J. , Toumani , F. : A Natural Language Mediation System for E-Commerce applications . In: Workshop HLT for the Semantic Web and Web Services. ISWC . ( 2003 ) 39 - 50

4. Ehrig , M. , Staab , S.: QOM - quick ontology mapping . In: ISWC . ( 2004 ) 683 - 697

5. Euzenat , J. , Shvaiko , P. : Ontology Matching. Springer, Heidelberg ( 2007 )

6. Heinecke , J. , Smits , G. , Chardenon , C. , Guimier De Neef , E. , Maillebuau , E. , Boualem , M.: TiLT : plate-forme pour le traitement automatique des langues naturelles . TAL 49:2 ( 2008 )

7. Giunchiglia , F. , Marchese , M. , Zaihrayeu , I. : Encoding classifications into lightweight ontologies . Journal of Data Semantics VIII ( 2007 ) 57 - 81

8. Reiter , N. , Hartung , M. , Frank , A. : A resource-poor approach for linking ontology classes to Wikipedia articles . In: STEP. ( 2008 ) 381 - 387

9. Chagnoux , M. , Heinecke , J.: Aligner ontologies et langues naturelles. ge´rer la synonymie . In: Plateforme AFIA , Grenoble ( 2007 ) 87 - 94