Weighted Transformer Network for Machine Translation

TL;DR

The paper introduces Weighted Transformer, an improved self-attention model for machine translation that converges faster and achieves higher BLEU scores than the baseline Transformer architecture.

Contribution

It proposes a novel weighted attention mechanism with multiple self-attention branches that enhances performance and training efficiency in neural machine translation.

Findings

Outperforms baseline Transformer in BLEU scores by 0.5 and 0.4 points on WMT tasks.

Converges 15-40% faster during training.

Achieves state-of-the-art results on translation benchmarks.

Abstract

State-of-the-art results on neural machine translation often use attentional sequence-to-sequence models with some form of convolution or recursion. Vaswani et al. (2017) propose a new architecture that avoids recurrence and convolution completely. Instead, it uses only self-attention and feed-forward layers. While the proposed architecture achieves state-of-the-art results on several machine translation tasks, it requires a large number of parameters and training iterations to converge. We propose Weighted Transformer, a Transformer with modified attention layers, that not only outperforms the baseline network in BLEU score but also converges 15-40% faster. Specifically, we replace the multi-head attention by multiple self-attention branches that the model learns to combine during the training process. Our model improves the state-of-the-art performance by 0.5 BLEU points on the WMT 2014 English-to-German translation task and by 0.4 on the English-to-French translation task.