Automatic Treebank Conversion via Informed Decoding

In this paper, we focus on the challenge of automatically converting a constituency treebank (source treebank) to fit the standard of another constituency treebank (target treebank). We formalize the conversion problem as an informed decoding procedure: information from original annotations in a source treebank is incorporated into the decoding phase of a parser trained on a target treebank during the parser assigning parse trees to sentences in the source treebank. Experiments on two Chinese treebanks show significant improvements in conversion accuracy over baseline systems, especially when training data used for building the parser is small in size.