Mining defining contexts to help structuring differential ontologies

In this paper, we present an experiment dealing with corpus-based construction of “differential ontologies”, which are organised according to semantic similarity and differential features. We argue that knowledge-rich defining contexts can be useful to help an ontology modeller in his task. We present a method, based on lexico-syntactic patterns, to spot such contexts in a corpus, then identify the terms they relate (definiendum and genus or “characteristics”) and the semantic relation that links them. We also show how potential co-hyponyms can be detected on the basis of shared words in their definiens. We evaluate the extracted defining sentences, semantic relations and co-hyponyms on a test corpus focusing on childhood and on an evaluation corpus about dietetics (both corpora are French). Definition extraction obtains 50% precision and recall of approximately 40%. Semantic relation identification reaches an average of 48% precision, and co-hyponyms 23.5%. We discuss the results of these experiments and conclude on perspectives for future work.