Encoding multi-granularity structural information for joint Chinese word segmentation and POS tagging

Abstract

Abstract Recent studies show that the joint Chinese word segmentation and POS tagging can enhance the mutual interaction and yield better performances for two tasks. However, existing joint methods fail to effectively take the advantage of the multiple granularity of information, e.g., character, word and subword, which has been proven prominently useful. In this paper, we propose to improve the joint tasks by leveraging such multi-granularity of information, by exploiting the lattice-LSTM and Convolutional Network (GCN) models for effectively encoding the graph information. On five benchmark datasets our proposed model shows highly competitive performances, achieving the new state-of-the-art results in the literature. Further analysis reveals that the multi-granularity information can relieve the out-of-vocabulary and the long-range dependency issues. Also the GCN structure is more effective for encoding the multi-granularity graph information, compared with the lattice structure.