Adding Interpretable Attention to Neural Translation Models Improves Word Alignment
This addresses the problem of interpretability in neural translation models for researchers and practitioners, though it is incremental as it builds on existing Transformer architectures.
The paper tackled the challenge of deriving accurate word alignments from state-of-the-art neural machine translation models by proposing a simple extension to the Transformer architecture and a novel alignment inference procedure, resulting in alignments that dramatically outperform naive methods and are comparable to Giza++ on two datasets.
Multi-layer models with multiple attention heads per layer provide superior translation quality compared to simpler and shallower models, but determining what source context is most relevant to each target word is more challenging as a result. Therefore, deriving high-accuracy word alignments from the activations of a state-of-the-art neural machine translation model is an open challenge. We propose a simple model extension to the Transformer architecture that makes use of its hidden representations and is restricted to attend solely on encoder information to predict the next word. It can be trained on bilingual data without word-alignment information. We further introduce a novel alignment inference procedure which applies stochastic gradient descent to directly optimize the attention activations towards a given target word. The resulting alignments dramatically outperform the naive approach to interpreting Transformer attention activations, and are comparable to Giza++ on two publicly available data sets.