Shared-Private Bilingual Word Embeddings for Neural Machine Translation

TL;DR

This paper introduces shared-private bilingual word embeddings for neural machine translation, improving performance and reducing parameters by enabling source and target embeddings to share features.

Contribution

The paper proposes a novel shared-private embedding approach that links source and target embeddings more closely and decreases model size in NMT.

Findings

Significant performance improvements over strong baselines.

Effective across 5 language pairs and diverse alphabets.

Reduces model parameters substantially.

Abstract

Word embedding is central to neural machine translation (NMT), which has attracted intensive research interest in recent years. In NMT, the source embedding plays the role of the entrance while the target embedding acts as the terminal. These layers occupy most of the model parameters for representation learning. Furthermore, they indirectly interface via a soft-attention mechanism, which makes them comparatively isolated. In this paper, we propose shared-private bilingual word embeddings, which give a closer relationship between the source and target embeddings, and which also reduce the number of model parameters. For similar source and target words, their embeddings tend to share a part of the features and they cooperatively learn these common representation units. Experiments on 5 language pairs belonging to 6 different language families and written in 5 different alphabets demonstrate that the proposed model provides a significant performance boost over the strong baselines with dramatically fewer model parameters.