A Sequence-to-Action Architecture for Character-Based Chinese Dependency Parsing with Status History

Character-based Chinese dependency parsing jointly learns Chinese word segmentation, POS tagging and dependency parsing to avoid the error propagation problem of pipeline models. Recent works on this task only rely on a local status for prediction at each step, which is insufficient for guiding global better decisions. In this paper, we first present a sequence-to-action model for character-based dependency parsing. In order to exploit decision history for prediction, our model tracks the status of parser particularly including decision history in the decoding procedure by employing a sequential LSTM. Additionally, for resolving the problem of high ambiguities in Chinese characters, we add position-based character embeddings to exploit character information with specific contexts accurately. We conduct experiments on Penn Chinese Treebank 5.1 (CTB-5) dataset, and the results show that our proposed model outperforms existing neural network system in dependency parsing, and performs preferable accuracy in Chinese word segmentation and POS tagging.