Computational Models of Similarity in Lexical Ontologies

This thesis mainly concerns itself with the issue of semantic similarity and computational applications of it. Semantic similarity has for a long time been a subject of intense scholarship in the fields of Artificial Intelligence, Psychology and Cognitive Science. Computational models trying to imitate aspects of this cognitive ability date back to Quillian and his spreading activation algorithm. In the research presented here we propose a new computational approach to the assessment of semantic similarity between concepts. This work builds upon previous information theoretic attempts, from researchers interested in natural language processing, to mimic this ill-defined human aptitude. WordNet, an ontological lexical knowledge base, is the only source of knowledge exploited in the proposed algorithm. Hence, this is where our work differs from previous information theoretic attempts, since these have relied upon statistical information gathered from analyzing semantically tagged corpora. As a consequence, the sparse data problem that these corpus based strategies have to face is completely avoided, as is the time consuming effort to construct tagged corpora. Experimental studies presented show that this new approach not only outperforms previous formulations in the information theoretic category, but also other non-information theoretic formulations. We also outline how an analogy can be established between information theory and set theory that allows us to view most information theoretic models in terms of Tversky’s feature model. As a consequence of the correspondence we also propose a new similarity measure based on Tversky’s model. We also study a specific application of similarity that, by way of polysemy, will allow us infer novel and useful re-categorizations of existing concepts in WordNet. These creative re-categorizations are inspired by the Torrance test of creative thinking. Since our proposal depends on the quality of the lexical ontology, we will also propose an algorithm that explores the conceptual space defined by the ontology and suggests the inclusion of new concepts. These new concepts can be added at specific locations in the ontology, filling-in some of the lexical holes and asymmetries existent in WordNet. The filling of these holes provide a better conceptual grounding on which to base our model of similarity.