Stance Detection with Bidirectional Conditional Encoding

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 876–885, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Stance Detection with Bidirectional Conditional Encoding Isabelle Augenstein and Tim Rockt¨aschel Department of Computer Science University College London i.augenstein@ucl.ac.uk, t.rocktaschel@cs.ucl.ac.uk Andreas Vlachos and Kalina Bontcheva Department of Computer Science University of Sheffield {a.vlachos, k.bontcheva}@sheffield.ac.uk Abstract Stance detection is the task of classifying the attitude Previous work has assumed that ei- ther the target is mentioned in the text or that training data for every target is given. This pa- per considers the more challenging version of this task, where targets are not always men- tioned and no training data is available for the test targets. We experiment with condi- tional LSTM encoding, which builds a rep- resentation of the tweet that is dependent on the target, and demonstrate that it outperforms encoding the tweet and the target indepen- dently. Performance is improved further when the conditional model is augmented with bidi- rectional encoding. We evaluate our approach on the SemEval 2016 Task 6 Twitter Stance Detection corpus achieving performance sec- ond best only to a system trained on semi- automatically labelled tweets for the test tar- get. When such weak supervision is added, our approach achieves state–of-the-art results. 1 Introduction The goal of stance detection is to classify the attitude expressed in a text towards a given target, as “pos- itive”, ”negative”, or ”neutral”. Such information can be useful for a variety of tasks, e.g. Mendoza et al. (2010) showed that tweets stating actual facts were affirmed by 90% of the tweets related to them, while tweets conveying false information were pre- dominantly questioned or denied. In this paper we focus on a novel stance detection task, namely tweet stance detection towards previously unseen targets (mostly entities such as politicians or issues of pub- lic interest), as defined in the SemEval Stance De- tection for Twitter task (Mohammad et al., 2016). This task is rather difficult, firstly due to not having training data for the targets in the test set, and sec- ondly, due to the targets not always being mentioned in the tweet. For example, the tweet “@realDon- aldTrump is the only honest voice of the @GOP” expresses a positive stance towards the target Don- ald Trump. However, when stance is annotated with respect to Hillary Clinton as the implicit target, this tweet expresses a negative stance, since supporting candidates from one party implies negative stance towards candidates from other parties. Thus the challenge is twofold. First, we need to learn a model that interprets the tweet stance towards a target that might not be mentioned in the tweet it- self. Second, we need to learn such a model without labelled training data for the target with respect to which we are predicting the stance. In the example above, we need to learn a model for Hillary Clinton by only using training data for other targets. While this renders the task more challenging, it is …

Content truncated to 500 words