Incorporating Stock Market Signals for Twitter Stance Detection

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4074 - 4091 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Incorporating Stock Market Signals for Twitter Stance Detection Costanza Conforti1, Jakob Berndt2, Mohammad Taher Pilehvar1,3, Chryssi Giannitsarou2, Flavio Toxvaerd2, Nigel Collier1 1 Language Technology Lab, University of Cambridge 2 Faculty of Economics, University of Cambridge 3 Tehran Institute for Advanced Studies, Khatam University, Iran {cc918,jb2088}@cam.ac.uk Abstract Research in stance detection has so far focused on models which leverage purely textual in- put. In this paper, we investigate the integra- tion of textual and financial signals for stance detection in the financial domain. Specifically, we propose a robust multi-task neural archi- tecture that combines textual input with high- frequency intra-day time series from stock market prices. Moreover, we extend WT–WT, an existing stance detection dataset which col- lects tweets discussing Mergers and Acquisi- tions operations, with the relevant financial sig- nal. Importantly, the obtained dataset aligns with STANDER, an existing news stance detec- tion dataset, thus resulting in a unique multi- modal, multi-genre stance detection resource. We show experimentally and through detailed result analysis that our stance detection sys- tem benefits from financial information, and achieves state-of-the-art results on the WT–WT dataset: this demonstrates that the combina- tion of multiple input signals is effective for cross-target stance detection, and opens inter- esting research directions for future work. 1 Introduction Stance detection (SD) is the task of automatically classifying the writer’s opinion expressed in a text towards a particular target (Küçük and Can, 2020). Starting from Mohammad et al. (2016)’s seminal work, research on Twitter SD gained increasing popularity (Ghosh et al., 2019), embracing new topics (Derczynski et al., 2017; Aker et al., 2017a; Conforti et al., 2020b) and languages (Gorrell et al., 2019; Vamvas and Sennrich, 2020a; Zotova et al., 2020). In recent years, research on SD has mainly focused on cross-target generalization, in which an SD system is tested on targets unseen during train- ing (Xu et al., 2018). Cross-target generalization constitutes one of the biggest challenges in Twitter SD (AlDayel and Magdy, 2021): in this context, researchers investigated a wide range of techniques, including adversarial training (Wang et al., 2020; Allaway et al.), cross-lingual transfer (Mohtarami et al., 2019), knowledge transfer using semantic and emotion lexicons (Zhang et al., 2020), weak supervision through synthetic samples (Conforti et al., 2021b; Li and Caragea, 2021), and various types of cross-domain transfer (Schiller et al., 2021; Hardalov et al., 2021a). In this paper, we study multimodality as a means to enhance cross-target generalization in Twitter SD. Multimodal Machine Learning studies the inte- gration and modeling of multiple modalities (Elliott et al., 2016), where a modality refers to the way in which something happens (Baltrusaitis et al., 2019). Our contributions are as follows: 1. We study multimodal learning for Twitter SD. Despite being an established research area in NLP (Elliott et al., 2016), SD in a multimodal context is still understudied. 2. We extend WT–WT, …

Content truncated to 500 words