[4Yin2] Interactive session 2

[4Yin2-40] Lexical Entailment with Hierarchy Representations by Deep Metric Learning

〇Naomi Sato¹, Masaru Isonuma¹, Kimitaka Asatani¹, Shoya Ishiduka², Aori Simizu², Ichiro Sakata¹ (1.The University of Tokyo, 2.Daikin Industries, Ltd.)

Keywords:Representation Learning, Lexical Entailment, Deep Metric Learning

A thesaurus synthesizes synonyms and hypernyms as a hierarchical structure, and it has been used in various tasks. To automatically construct a thesaurus, a variety of lexical entailment methods have been proposed, including post-processing of distributed representations and embedding hierarchical structures into hyperbolic spaces. However, their methods have many issues for practical application, such as low generalization performance for words not included in the thesaurus and difficulty in designing tasks on hyperbolic spaces. This study proposes a method that learns a mapping from words to the hierarchical embedding representation by using Ranked List Loss, and it enables embedding any input words into the Cartesian coordinate system. We evaluated the performance on the lexical entailment tasks and confirmed that the proposed method achieves higher accuracy and generalization performance than existing methods.