Query-Key Normalization for Transformers
This addresses the problem of improving translation quality for low-resource languages, which is socially valuable but challenging, though it appears incremental as it builds on existing normalization adaptations.
The paper tackled low-resource language translation by proposing QKNorm, a normalization technique for Transformers that modifies the attention mechanism to reduce softmax saturation, resulting in an average improvement of 0.928 BLEU over state-of-the-art benchmarks on 5 low-resource translation pairs.
Low-resource language translation is a challenging but socially valuable NLP task. Building on recent work adapting the Transformer's normalization to this setting, we propose QKNorm, a normalization technique that modifies the attention mechanism to make the softmax function less prone to arbitrary saturation without sacrificing expressivity. Specifically, we apply $\ell_2$ normalization along the head dimension of each query and key matrix prior to multiplying them and then scale up by a learnable parameter instead of dividing by the square root of the embedding dimension. We show improvements averaging 0.928 BLEU over state-of-the-art bilingual benchmarks for 5 low-resource translation pairs from the TED Talks corpus and IWSLT'15.