A Siamese Neural Network Framework With Sememe-Based Context Extraction for Interactive Argument Pair Identification

Ning Yu, Jianyi Liu, Yu Shi

Abstract: Interactiveargument pair identification has become an essential task in natural language processing. Existing work points out that key text segments in the context of an argument are crucial for identification performance. However, current context extraction methods ignore various commonsense knowledge implicit in arguments and contexts, which are very important for deep semantic mining of text and semantic similarity calculation. Furthermore, few methods consider the relation between argument pairs with different interactivity. In this letter, a sememe-based context extraction module is proposed to solve the first problem. The module can use the knowledge in HowNet to calculate the deep semantic similarity and further extract strongly relevant text segments in the context of the argument. We also propose a siamese neural network framework to solve the second problem. It enables two arguments to interact fully, models the relation between two argument pairs, and introduces the triplet loss for optimization. Experimental results show that the proposed method outperforms the state-of-the-art method in the benchmark dataset. Further analysis demonstrates the effectiveness of our proposed method.