Coreference Resolution with Hierarchical Pointer Networks Based on Pointing Methods

Coreference resolution is a type of discourse analysis task of natural language processing. The task entails a method of linking different words expressed in an arbitrary entity within a document. Pointer networks (Ptr-net) based on RNN encoder-decoder is capable of modeling the problem of a target class variable and learning and predicting a position corresponding to a given input sequence. A hierarchical RNN encoder-decoder is an extended model of the RNN encoder-decoder that performs encoding on a word-by-word basis for a given input sequence. It performs encoding on a sentence basis to predict the output result. Ptr-net suffers from performance degradation when the input is composed of multiple sentences, or when the length of the input sentence is long. To solve this problem, we propose a coreference resolution with a hierarchical pointer networks (HRPT) model based on pointing methods for all cases mentioned. The HRPT encodes word and sentence levels for the input sequence, consisting of multiple sentences. It uses both word and sentence-level information in the decoder. The pointing method for coreference resolution is a way of pointing to the position of the referenced word or the referenced pivot of the entity. Experimental results show that the proposed model is CoNLL F1 71.56%, which is 20.96% better than the rule-based model.