Rbosaghz writeup of Mike Collins Talk

Mike_Collins_Talk by user:Rbosaghz

The main thrust of this talk was using TAG-style parse trees to aid Phrase-based Statistical Machine Translation. Prof. Collins started with an explanation of Parse trees, quickly moving away from Context Free Grammars to Tree Adjoining Grammars (TAG). After describing TAG, he outlined how one can use CRFs to parse TAG. From there he moved onto using TAG trees for Machine Translation.

After describing a vanilla phrase-based machine translation system (translation model + language model), he described how one can use TAG trees to aid the translation pipeline. He compared their system to Pharoah, a popular phrase-based MT system.

I liked the new ideas and use of CRFs for parsing in this paper, however, the gains in BLEU score were very weak (less than 1%). Furthermore, the baseline Pharoah system is a very old and outdated baseline to compare against. I would go as far as to say that under the BLEU score, TAG probably doesn't help. But the ideas are nice and worth discussing.