Stance Detection with Hierarchical Attention Network

Proceedings of the 27th International Conference on Computational Linguistics, pages 2399–2409 Santa Fe, New Mexico, USA, August 20-26, 2018. 2399 Stance Detection with Hierarchical Attention Network Qingying Sun1,3, Zhongqing Wang1, Qiaoming Zhu1,2 and Guodong Zhou1 1School of Computer Science and Technology, Soochow University, Suzhou, China 2Institute of Artificial Intelligence, Soochow University, Suzhou, China 3Huaiyin Normal University, Huaian, China qingying.sun@foxmail.com, wangzq.antony@gmail.com, {qmzhu, gdzhou}@suda.edu.cn Abstract Stance detection aims to assign a stance label (for or against) to a post toward a specific target. Recently, there is a growing interest in using neural models to detect stance of documents. Most of these works model the sequence of words to learn document representation. However, much linguistic information, such as polarity and arguments of the document, is correlated with the stance of the document, and can inspire us to explore the stance. Hence, we present a neural model to fully employ various linguistic information to construct the document representation. In addition, since the influences of different linguistic information are different, we propose a hi- erarchical attention network to weigh the importance of various linguistic information, and learn the mutual attention between the document and the linguistic information. The experimental re- sults on two datasets demonstrate the effectiveness of the proposed hierarchical attention neural model. 1 Introduction While a lot of works on document-level opinion mining have involved determining the polarity expressed in a customer review (Pang et al., 2008; Liu, 2012), researchers have begun exploring new opinion mining tasks in recent years. One such task is stance detection: given a post written for a two-sided topic discussed in an online debate forum, determine which of the two sides (i.e., for and against) its author is taking. Stance detection system can potentially have a positive social impact and are of practical interest to non-profits, governmental organizations, and companies. Previously, some of the related researches used feature engineering to extract either linguistic (Soma- sundaran and Wiebe, 2010) or structure (Sridhar et al., 2015) features manually. Recently, neural models have achieved high success and obtained the best results on stance detection (Zarrella and Marsh, 2016; Du et al., 2017). A key issue of the neural model is document representation, and most of previous works learn docu- ment representation from word sequence using Convolutional Neural Networks (CNNs) (Chen and Ku, 2016; Vijayaraghavan et al., 2016) or Recurrent Neural Networks (RNNs) (Augenstein et al., 2016; Du et al., 2017). E1: I understand that homosexuals can't have kids. They can adopt children and thus, support the advancement of the human race. (Target: Gay Rights) nsubj dobj dobj nmod:of Arguments pairs: (adopt,they),(adopt,children),(support,advancement),(advancement,race) However, besides the word sequence, stance is correlated with some explicit and implicit linguistic factors. Take E1 for example, both sentimental words (i.e., understand, support) and argument sentence (i.e., They can adopt ...) support the favor stance toward the target “Gay Rights”. In addition, the dependency pair, (adopt, children) and (advancement, race), also express a positive stance. Hence, This work is licensed under a Creative Commons Attribution 4.0 International License. …

Content truncated to 500 words