Is it possible align phrase by phrase, not just word by word?

[fishfree]

For many languages, there are lots of unpaired words, but lots of paired phrase.

[zdou0830]

It's possible as here we first compute subword alignments and then convert them to word alignments, thus we can use the same strategy to convert them to phrase alignments.

[lofti198]

For many languages, there are lots of unpaired words, but lots of paired phrase.

Which language pair you use?

[fishfree]

@lofti198 For example, English & Chinese.

@zdou0830 Thank you. Then need we manually annotate the gold prase alignment corpus first?

[zdou0830]

You'll need to do phrase segmentation and use this segmentation during inference. Changing our word-level self-training objective (detail in the paper) to phrase-level self-training objective may also be helpful. If you have gold phrase alignment information, you can also use our model in supervised settings (https://github.com/neulab/awesome-align#supervised-settings) to train the model.