Sense Vocabulary Compression through the Semantic Knowledge of WordNet for Neural Word Sense Disambiguation

TL;DR

This paper introduces methods to compress WordNet's sense vocabulary using semantic relationships, reducing model complexity and improving coverage in neural word sense disambiguation without extra data, while leveraging BERT for superior performance.

Contribution

It proposes two novel sense vocabulary compression techniques that enhance neural WSD models' efficiency and coverage, combined with a BERT-based system that outperforms existing methods.

Findings

Reduced sense vocabulary size improves model efficiency.

Enhanced coverage without additional training data.

Achieved state-of-the-art results on WSD benchmarks.

Abstract

In this article, we tackle the issue of the limited quantity of manually sense annotated corpora for the task of word sense disambiguation, by exploiting the semantic relationships between senses such as synonymy, hypernymy and hyponymy, in order to compress the sense vocabulary of Princeton WordNet, and thus reduce the number of different sense tags that must be observed to disambiguate all words of the lexical database. We propose two different methods that greatly reduces the size of neural WSD models, with the benefit of improving their coverage without additional training data, and without impacting their precision. In addition to our method, we present a WSD system which relies on pre-trained BERT word vectors in order to achieve results that significantly outperform the state of the art on all WSD evaluation tasks.