Two-Target Stance Detection with Target-Related Zone Modeling

Recently, stance detection on Twitter has been a hot research topic in the natural language processing community. Most previous studies have assumed that only one target is involved in a tweet and perform single-target stance detection. In this study, we address a more challenging version of this task, namely two-target stance detection, where two targets are involved. Specifically, we first define four categories of text zones related to two targets, i. e., Target 1, Target 2, Non-target, and Other, and propose an unsupervised approach to automatically obtain these zones. Then, we propose a hierarchical neural network to perform stance detection for each target where multiple LSTM layers are leveraged to encode the target-related zones and a Bidirectional LSTM layer to encode the outputs from the LSTM layers. Moreover, we introduce a target-related attention mechanism in the hierarchical network. Empirical studies demonstrate the effectiveness of the proposed approach to two-target stance detection.