Event Coreference Fusion Based on Deep Residual Network and Structure Representation

Abstract

Event coreference fusion is a challenging task in natural language processing, which plays an important role in event extraction, question answering system and reading comprehension. In this paper, a deep residual network model Res-HASR with hierarchical attention and event structure representation is proposed for event coreference fusion. Res- HASR first uses bi-directional long-short-term memory(Bi- LSTM)network and hierarchical attention to obtain the important context information of event sentences; and deep residual network (ResNET) is used to obtain the deep semantic information of event sentences; then uses neural tensor network to calculate the co-occurrence relationship of entities and elements in event sentence and structure representation; then designs the element fusion algorithm to fuse the elements; and finally the events after fusion is used iteratively for event coreference resolution to further improve system performance. The experimental results illustrate that Res-HASR improves the performance of event coreference fusion significantly compared with the two basic models.