Synonyms, Antonyms and Beyond

Abstract

This paper examines fine-tuning for a well-studied task, synonym/antonym classification, and a novel task, VAD regression. The proposed method is almost as good as SOTA methods on standard benchmarks, but generalizes better to a new benchmark based on the Fallows Thesaurus (1898). Moreover the proposed method generalizes from words to (multilingual) texts: e.g., MWEs, OOVs, sentences, paragraphs. The second task was introduced to address concerns with leakage in benchmarks. The second task also makes it easier to study the (in)-effectiveness of transfer learning to unseen words.

Tasks

Reproductions