Tweet Stance Detection Using an Attention based Neural Ensemble Model

Proceedings of NAACL-HLT 2019, pages 1868–1873 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 1868 Tweet Stance Detection Using an Attention based Neural Ensemble Model Umme Aymun Siddiqua, Abu Nowshed Chy, and Masaki Aono Department of Computer Science and Engineering Toyohashi University of Technology, Toyohashi, Aichi, Japan {aymun,nowshed}@kde.cs.tut.ac.jp, and aono@tut.jp Abstract Stance detection in twitter aims at mining user stances expressed in a tweet towards a single or multiple target entities. To tackle this problem, most of the prior studies have been explored the traditional deep learning models, e.g., LSTM and GRU. However, in compared to these traditional approaches, re- cently proposed densely connected Bi-LSTM and nested LSTMs architectures effectively address the vanishing-gradient and overfitting problems as well as dealing with long-term de- pendencies. In this paper, we propose a neu- ral ensemble model that adopts the strengths of these two LSTM variants to learn better long-term dependencies, where each module coupled with an attention mechanism that am- plifies the contribution of important elements in the final representation. We also employ a multi-kernel convolution on top of them to extract the higher-level tweet representa- tions. Results of extensive experiments on sin- gle and multi-target stance detection datasets show that our proposed method achieves sub- stantial improvement over the current state-of- the-art deep learning based methods. 1 Introduction Tweet stance detection is the task of automati- cally determining the stance of a tweet whether the tweet is in favor of, against, or none towards a tar- get (Mohammad et al., 2017). We can consider it as a sub-domain of sentiment analysis. However, the goal of sentiment analysis is to classify the po- larity of a tweet sentiment based on its contents, whereas identification of stance is dependent on the specific target. For example, Figure 1 depicts the stance of a tweet towards different targets. Stance detection in twitter poses unique chal- lenges to the research community since tweets are short and informal user-generated text, which usually tend not to follow the grammatical rules. Sample Tweet: @realDonaldTrump Crazy or blind, death and mute? You liberal troll. #TRUMP IS GOING TO BEAT #Hillary LANDSLIDE. #TrumpTsunami Target II: Donald Trump Stance II: Favor Target I: Hillary Clinton Stance I: Against Figure 1: Example of target-specific stance detection. Moreover, tweets contain plenty of idiosyncratic abbreviations as well as other twitter specific syn- taxes such as #hashtags and emoticons. To ad- dress the challenges of stance detection in twit- ter, Mohammad et al. (2016) presented a tweet stance detection task that focused on a single tar- get in SemEval-2016. Top performing systems in this task proposed several deep learning based approaches by using CNN (Wei et al., 2016), RNN (Zarrella and Marsh, 2016), and so on. Later, (Du et al., 2017) utilized the target- augmented embeddings in an attention based neu- ral network, whereas (Zhou et al., 2017) proposed an attention mechanism at the semantic level in the bidirectional GRU-CNN structure to perform target-specific stance detection on tweets. More recently, (Dey …

Content truncated to 500 words