Unsupervised learning of an IS-A taxonomy from a limited domain-specific corpus

This report addresses the problem of learning a taxonomy from a given domain-specific text corpus. We propose a novel unsupervised algorithm for this problem. Its key contributions include a clustering-based inference approach that increases recall over surface patterns and a graph-based algorithm for detecting incorrect edges that improves precision. Our system induces the taxonomy simply by analyzing the provided corpus. Thus, the learned taxonomy is focused on the concepts that are relevant for the specific corpus. An empirical evaluation on five corpora demonstrates the utility of the system.