Tweet Stance Detection: A Two-stage DC-BILSTM Model Based on Semantic Attention

Abstract

Stance classification in tweet aims at detecting whether the author of the tweet is in FAVOR of, AGAINST, or NONE towards a pre-chosen target entity. Recently proposed Densely Connected BI-LSTM can effectively relieve overfitting and vanishing-gradient problems as well as dealing with long-term dependencies during multi-layer LSTM training. Based on this, we propose a two-stage deep attention neural network(T-DAN) for target-specific stance detection. This model employs densely connected BI-LSTM to encode tweet tokens and traditional bidirectional LSTM to encode target tokens. Besides, we decompose this ternary classification problem into two binary classification problems to mitigating the imbalanced distribution of labels. In the first stage, we find out the tweet is neutral or subjective about the specific target. In the second stage, we classify the stance of a given subjective tweet’s stance. Moreover, we propose a novel method of attention calculation based on the semantic similarity of tweet tokens and target tokens which can locate the crucial words related to target. Experimental results on English and Chinese datasets demonstrate that our proposed method surpasses some strong baselines and achieves the stateof-the-art performance.