Combining Subword Representations into Word-level Representations in the Transformer Architecture

In Neural Machine Translation, using word-level tokens leads to degradation in translation quality. The dominant approaches use subword-level tokens, but this increases the length of the sequences and makes it difficult to profit from word-level information such as POS tags or semantic dependencies. We propose a modification to the Transformer model to combine subword-level representations into word-level ones in the first layers of the encoder, reducing the effective length of the sequences in the following layers and providing a natural point to incorporate extra word-level information. Our experiments show that this approach maintains the translation quality with respect to the normal Transformer model when no extra word-level information is injected and that it is superior to the currently dominant method for incorporating word-level source language information to models based on subword-level vocabularies.