Extraction of Lexico-Semantic Classes from Text

This paper describes an unsupervised method for extracting lexico-semantic classes from POS annotated corpora. The method consists in building bi-dimensional clusters of both words and local syntactic contexts. Each cluster, which represents a lexicosemantic class such as “entities in danger” is the result of merging its most prototypical constituents (words and contexts). The generated clusters will be used as centroids to word classification. The basic intuition underlying our corpus-based approach is that similar classes can be aggregated to generate either more specific or more generic classes, without inducing odd associations between contexts and words. A new class is generated by specification if we make the union of the constituent contexts (intension expansion) while the words are intersected (intension reduction). A new class is generated by abstraction if the local contexts are intersected (intension reduction), while we make the union of the constituent words (extension expansion). Intersecting words and local contexts in an accurate way allows us to generate tight clusters with prototypical constituents.