Enhancing Cross-target Stance Detection with Transferable Semantic-Emotion Knowledge

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3188–3197 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 3188 Enhancing Cross-target Stance Detection with Transferable Semantic-Emotion Knowledge Bowen Zhang1, Min Yang2, Xutao Li3∗, Yunming Ye3∗, Xiaofei Xu1, Kuai Dai3 1Harbin Institute of Technology, Harbin, China 2Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, China 3Harbin Institute of Technology, Shenzhen, China Abstract Stance detection is an important task, which aims to classify the attitude of an opinionated text towards a given target. Remarkable suc- cess has been achieved when sufficient labeled training data is available. However, annotating sufficient data is labor-intensive, which estab- lishes significant barriers for generalizing the stance classifier to the data with new targets. In this paper, we proposed a Semantic-Emotion Knowledge Transferring (SEKT) model for cross-target stance detection, which uses the external knowledge (semantic and emotion lexicons) as a bridge to enable knowledge transfer across different targets. Specifically, a semantic-emotion heterogeneous graph is con- structed from external semantic and emotion lexicons, which is then fed into a graph convo- lutional network to learn multi-hop semantic connections between words and emotion tags. Then, the learned semantic-emotion graph rep- resentation, which serves as prior knowledge bridging the gap between the source and tar- get domains, is fully integrated into the bidi- rectional long short-term memory (BiLSTM) stance classifier by adding a novel knowledge- aware memory unit to the BiLSTM cell. Exten- sive experiments on a large real-world dataset demonstrate the superiority of SEKT against the state-of-the-art baseline methods. 1 Introduction The goal of stance detection is to automatically predict the attitude (i.e., favor, against, or none) of an opinionated text towards a given target (Du et al., 2017). Recently, deep learning methods, such as convolutional neural network (CNN) and long short-term memory (LSTM) (Augenstein et al., 2016; Du et al., 2017), have dominated the study of stance detection. Impressive stance detection performances have been achieved when a large ∗corresponding authors: {lixutao, yym}@hit.edu.cn number of labeled samples are available. However, obtaining rich annotated data is a time-consuming and labor-intensive process. Conventional stance detection methods are struggling to cope well with the data across targets. This motivates the stud- ies of cross-target stance detection (Wei and Mao, 2019), which infers the attitude of the destination target by leveraging a large amount of annotated data from the source target. So far, several previous studies have been con- ducted for cross-target stance detection (Augen- stein et al., 2016; Xu et al., 2018; Wei and Mao, 2019). These methods leverage either common words or concept-level knowledge shared by dif- ferent targets to bridge the knowledge gap across the different targets. Such models suffer from two issues when they are applied to cross-target stance detection in practice. First, stance detection often involves analyzing the texts from social media that are short and informal, making it difficult to ex- tract domain-independent common words shared by different targets from the training data. Second, users may express their stance towards a …

Content truncated to 500 words