Encoders Help You Disambiguate Word Senses in Neural Machine Translation
This addresses the challenge of improving translation accuracy for ambiguous words in NMT, which is an incremental contribution to model interpretability.
The paper tackled the problem of understanding which components in neural machine translation (NMT) dominate word sense disambiguation, finding that encoder hidden states significantly outperform word embeddings in predicting correct translations for ambiguous nouns, with decoders providing additional relevant information.
Neural machine translation (NMT) has achieved new state-of-the-art performance in translating ambiguous words. However, it is still unclear which component dominates the process of disambiguation. In this paper, we explore the ability of NMT encoders and decoders to disambiguate word senses by evaluating hidden states and investigating the distributions of self-attention. We train a classifier to predict whether a translation is correct given the representation of an ambiguous noun. We find that encoder hidden states outperform word embeddings significantly which indicates that encoders adequately encode relevant information for disambiguation into hidden states. Decoders could provide further relevant information for disambiguation. Moreover, the attention weights and attention entropy show that self-attention can detect ambiguous nouns and distribute more attention to the context. Note that this is a revised version. The content related to decoder hidden states has been updated.