Home

Welcome to the WSD wiki!

Introduction

In this work we present a Word Sense Disambiguation (WSD) engine that integrates a Transformer-based neural architecture with knowledge present in WordNet, the resource from which the sense inventory is taken from.

Model

The architecture is composed of ELMo embeddings plus a TransformerXL (x3) on top with a final dense layer for tagging each word with the right lemma, pos, and sense identifier.

To incorporate lexical knowledge at evaluation time where we score each possible sense of a word with different scores:

• the semantic similarity of the context with the gloss of the sense and it's direct hypernyms and hyponyms.
• the accumulated probabilities of BERT language model for the lemma names of the synset and of its direct hypernyms and hyponyms.

Data Source

The training and test data was taken from http://lcl.uniroma1.it/wsdeval/

References Links

• ELMo: https://arxiv.org/abs/1802.05365
• Universal Sentence Encoder: https://arxiv.org/abs/1803.11175
• GAS: https://aclweb.org/anthology/P18-1230
• BERT: https://arxiv.org/abs/1810.04805
• WSD Evaluation Framework: http://wwwusers.di.uniroma1.it/~navigli/pubs/EACL_2017_Raganatoetal.pdf
• Sentence encoding survey: https://arxiv.org/abs/1806.06259