Acquiring Human-like Feature-Based Conceptual Representations from Corpora

Abstract

The automatic acquisition of feature-based conceptual representations from text corpora can be challenging, given the unconstrained nature of human-generated features. We examine large-scale extraction of concept-relation-feature triples and the utility of syntactic, semantic, and encyclopedic information in guiding this complex task. Methods traditionally employed do not investigate the full range of triples occurring in human-generated norms (e.g. flute produce sound), rather targeting concept-feature pairs (e.g. flute - sound) or triples involving specific relations (e.g. is-a, part-of). We introduce a novel method that extracts candidate triples (e.g. deer have antlers, flute produce sound) from parsed data and re-ranks them using semantic information. We apply this technique to Wikipedia and the British National Corpus and assess its accuracy in a variety of ways. Our work demonstrates the utility of external knowledge in guiding feature extraction, and suggests a number of avenues for future work.