Generalization of Words for Chinese Dependency Parsing

In this paper, we investigate the influence of generalization of words to the accuracies of Chinese dependency parsing. Specially, in our shift-reduce parser, we use a neural language model based word embedding (NLMWE) method (Bengio et al., 2003) to generate distributed word feature vectors and then perform K-means based word clustering to generate word classes. We designed feature templates by making use of words, part-of-speech (POS) tags, coarse-grained POS (CPOS) tags, NLMWE-based word classes and their combinations. NLMWE-based word classes is shown to be an important supplement of POS-tags, especially when POS-tags are automatically generated. Experiments on a Query treebank, CTB5 and CTB7 show that the combinations of features from CPOS-tags, POS-tags, and NLMWE-based word classes yield the best unlabelled attachment scores (UASs). Our final UAS−p (excluding punctuations) of 86.79% on the CTB5 test set is comparable to state-of-theart results. Our final UAS−p of 86.80% and 87.05% on the CTB7 Stanford dependency test set and original test set is significantly better than three well known open-source dependency parsers.