Explaining Similarity of Terms

Computing the similarity between entities is a core component of many NLP tasks such as measuring the semantic similarity of terms for generating a distributional thesaurus. In this paper, we study the problem of explaining post-hoc why a set of terms are similar. Given a set of terms, our task is to generate a small set of explanations that best characterizes the similarity of those terms. Our contributions include: 1) an information-theoretic objective function for quantifying the utility of an explanation set; 2) a survey of psycholinguistics and philosophy for evidence of different sources of explanations such as descriptive properties and prototypes; 3) computational baseline models for automatically generating various types of explanations; and 4) a qualitative evaluation of our explanation generation engine.

• Type: Conference
• Year: 2008
• Where: COLING
• Authors: Vishnu Vyas, Patrick Pantel
• Keywords: COLING 2008, Computational Linguistics, Distributional Thesaurus, Information-theoretic Objective Function, Semantic Similarity