Multi-Target Stance Detection with Multi-Task Learning

Abstract

Stance detection plays an important role in the field of online public opinion analysis. The majority of methods of stance detection are for a single target at a once. However, in some situations, multi-target to be analyzed are interrelated, such as several different candidates in an election and different brands of the same product. Detecting the stance of a single target in isolation may lose some information. Some studies have shown that using multi-task learning mechanism for multi-target performs better results than only using a single task. Considering the effectiveness of sentiment features for stance detection in existing researches, we propose a multi-task learning model with sentiment features and at the same time use background lexicon to guide the attention mechanism. We reduce the work of manual labeling and use automatic methods to obtain background knowledge and sentiment information. We use BiLSTM to extract the semantic features and use Multi-Kernel Convolution to get the local features. We use the widely accepted evaluation method and our experiments achieve state-of-the-art results on a benchmark dataset.