Implicit discourse relation identification based on tree structure neural network

Ruiying Geng, Ping Jian, Yingxue Zhang, Heyan Huang

Abstract: This paper presents a tree structure neural network to predict the sense of implicit discourse relation in English. Tree structure neural network, also called recursive neural network, has been proved to be powerful in modeling compositionality in natural language using parse tree based structural representations. By integrating the semantic information along the parse trees, the tree structure neural network shows its superiority in capturing the structure information and composition semantics, which is meaningful for the deep semantic problems, such as sentiment analysis and discourse structure understanding. Experimental results obtained on Penn Discourse Tree Bank show that the tree structure neural network is more effective to predict the logical semantic relations between discourse texts when compared with the traditional shallow feature classifiers and sequential deep semantic models.