Multiple Feature-Sets Method for Dependency Parsing

This paper presents a simple and effective approach to improve dependency parsing by exploiting multiple feature-sets. Traditionally, features are extracted by applying the feature templates to all the word pairs(first-order features)and word tuples(second-order features). In this pa per, we show that exploiting different feature templates for different word pairs and word tuples achieves significant improvement overbaseline parsers. First, we train a text chunker using a freely available implementation of the first-order linear conditional random fields model. Then we build a clause-chunk tree for a given sentence based on chunking information and punctuation marks. Finally, we extract features for dependency parsing according to multiple feature-sets. We extend the projective parsing algorithms of McDonald[20] and Carreras[1] for our case, experimental results show that our approach significantly outperform the baseline systems without increasing complexity. Given correct chunking information, we improve from baseline accuracies of 91.36% and 92.20% to 93.19% and 93.89%, respectively.