Exploiting Lexical Relations For Semi-supervised Dependency Parsing

Dependency parsing plays an important role in Natural Language Processing.The target of this paper is to exploit lexical relations for dependency parsing.We conduct our research on two scenarios: monolingual and cross-lingual dependency parsing.For monolingual dependency parsing,we propose an approach to extracting coordinate word pairs to improve dependency parsing.In the approach,we first extract coordinate word pairs to build a coordinate word pair dictionary.Then,based on the coordinate word pair dictionary,we present a set of new features for the dependency parsing models.As for the task of cross-lingual dependency parsing,our target is to improve parsing performance on resource-poor languages which lack of human labeled data.In our approach,we automatically build labeled training data for target languages by knowledge transferring technologies.We transfer dependency parsing information from the source language to the target language based on word alignments.To improve parsing performance,we make use of additional lexical features and partially labeled sentences.Additional lexical features are composed of subtree-based features and word cluster-based features.Subtrees are extracted from auto labeled sentences and word clusters are obtained from monolingual unlabeled data.Finally,we design new features for subtrees and word clusters.On the other side,we focus on model training with partially labeled sentences.To make full use of partially labeled sentences,we change a partially labeled sentence into a dependency forest.In this way,we can train our dependency parsers based on partially labeled data directly.Moreover,we propose a POS tagging and dependency parsing joint model which support training on partially labeled sentences.The experimental results show that our proposed approach can significantly improve parsing performance on monolingual and multilingual datasets.