Stance Detection with Target and Target Towards Attention

Abstract

We propose a Neural Stance Detection model with target and target towards attention mechanism. Stance detection is the task of classifying the attitude towards a given target. Even though a variety of recurrent neural networks have been used in stance detection problems, existing modes only take advantage of target information and ignore target towards information. What's more, these models tend to perform well when the text discusses the target explicitly. However, when the target is implicitly mentioned, these models are not good. To address this problem, we introduce Target and Target towards Attention mechanism which takes not only target but also target towards information into account. This paper considers the more challenging version of this task, where targets are not always mentioned and a specific test target has no training data available. Our model first builds a hierarchical Long Short Term Memory (LSTM)[1] model to represent sentence and text. And then, target and target towards information are considered via attention mechanism over different semantic levels. We conduct our experiment on SemEval-2016 Task 6 dataset. And the results show that our model outperforms several strong baselines.