Neural Machine Translation with Synchronous Latent Phrase Structure

It has been reported that grammatical information is useful for machine translation (MT) tasks. However, the annotation of incurs significant human costs. Furthermore, it not trivial to adapt MT because usually employs tokenization standards might capture relation between two languages and consequently, subword such as byte-pair-encoding used alleviate out-of-vocabulary problems; however, this be compatible with those annotations. In work, we introduce methods incorporate without supervising explicitly: first, latent phrase structure induced in an unsupervised fashion from attention mechanism; second, structures encoder decoder are synchronized so they each other using constraints during training. We demonstrate our approach performs better tasks: word alignment, extra resources. found enhance precision alignments through synchronization constraint after exact alignment analysis.