Improving dependency parsers with supertags

Transition-based dependency parsing sys-tems can utilize rich feature representa-tions. However, in practice, features are generally limited to combinations of lexi-cal tokens and part-of-speech tags. In this paper, we investigate richer features based on supertags, which represent lexical tem-plates extracted from dependency struc-ture annotated corpus. First, we develop two types of supertags that encode infor-mation about head position and depen-dency relations in different levels of granu-larity. Then, we propose a transition-based dependency parser that incorporates the predictions from a CRF-based supertagger as new features. On standard English Penn Treebank corpus, we show that our su-pertag features achieve parsing improve-ments of 1.3 % in unlabeled attachment, 2.07 % root attachment, and 3.94 % in com-plete tree accuracy.