Taxonomy completion via implicit concept insertion

Abstract

High quality taxonomies play a critical role in various domains such as e-commerce, web search and ontology engineering. While there has been extensive work on expanding taxonomies from externally mined data, there has been less attention paid to enriching taxonomies by exploiting existing concepts and structure within the taxonomy. In this work, we show the usefulness of this kind of enrichment, and explore its viability with a new taxonomy completion system ICON (Implicit CONcept Insertion). ICON generates new concepts by identifying implicit concepts based on the existing concept structure, generating names for such concepts and inserting them in appropriate positions within the taxonomy. ICON integrates techniques from entity retrieval, text summary, and subsumption prediction; this modular architecture offers high flexibility while achieving state-of-the-art performance. We have evaluated ICON on two e-commerce taxonomies, and the results show that it offers significant advantages over strong baselines including recent taxonomy completion models and the large language model, ChatGPT.