Improving Anaphora Resolution by Animacy Identification
Yuanqing Zhu, Wei Song, Xianjun Liu, Lizhen Liu, Xinlei Shelly Zhao · 2019
Identification of entities' attributes plays a crucial role in anaphora resolution which will influence the final resolution result. This paper focuses on animacy identification and propose a method based on sequence label and it can improve anaphora resolution's performance effectively. This method built a bidirectional LSTM network to capture the information contained in the long text sequence effectively. Taking the output of the network as an intermediate result, then input it to the CRF structure, the animacy identification model was obtained through the organic combination of the two methods. We design two experiments in Chinese corpus. Experiments show that this method can effectively improve the performance of anaphora resolution.