Definition grounding for Taxonomy construction

Abstract

Taxonomy construction consists in building a tree of parent-child entities, specifically, hypernym hyponym pairs. When tackling Natural Language Processing (NLP) tasks with a Large Language Model (LLM), we may rely on the broad model’s knowledge derived from its training data, typically a large text collection obtained from public internet crawls. However the understanding of relationships between objects and concepts in a domain specific setting often differ from common knowledge or common sense. Using small scale experiments, we show that defining each taxon in context makes constructing specialized zoological taxonomies reliable.

Keywords