Sgopal1 writeup on collins talk

In this talk, Collins gave an overview of TAG based parsing and how it can be used in MT. The first part of the talk went into the details about Tree adjoining Grammars and related examples etc. The second part of the talk was focused on how it is applied to machine translation.

• In TAG, there is a set of spines - elementary line trees. There are operations defined on the spines. Applying those operations on the spines repeatedly gives rise to some sort of structure ( dependencies ) between spines.

• A conditional log-linear model is used to derive the parse tree. The feature are based on the spines and dependencies between them.

• He then went on to talk about how TAG is used in MT to capture the correspondence between the original sentence and its translation. TAG allows a richer set of combination of the spines, thereby any permutation of spines between the original and translated version can be captured.

• There is .5% improvement in BLEU score.