SSSD: Leveraging Pre-trained Models and Semantic Search for Semi-supervised Stance Detection

Proceedings of Recent Advances in Natural Language Processing, pages 264–273 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_030 264 SSSD: Leveraging Pre-Trained Models and Semantic Search for Semi-Supervised Stance Detection Andr´e Mediote de Sousa Informatics Institute Federal Univ. of Rio Grande do Sul Porto Alegre - Brazil andremediote@inf.ufrgs.br Karin Becker Informatics Institute Federal Univ. of Rio Grande do Sul Porto Alegre - Brazil karin.becker@inf.ufrgs.br Abstract Pre-trained models (PTMs) based on the Trans- formers architecture are trained on massive amounts of data and can capture nuances and complexities in linguistic expressions, making them a powerful tool for many natural language processing tasks. In this paper, we present SSSD (Semantic Similarity Stance Detection), a semi-supervised method for stance detection on Twitter that automatically labels a large, domain-related corpus for training a stance clas- sification model. The method assumes as input a domain set of tweets about a given target and a labeled query set of tweets of representative arguments related to the stances. It scales the automatic labeling of a large number of tweets, and improves classification accuracy by lever- aging the power of PTMs and semantic search to capture context and meaning. We largely out- performed all baselines in experiments using the Semeval benchmark. 1 Introduction Stance Detection (SD) is the task that automatically determines whether the author of a text is in favor of, against or does not manifest about a given target. Targets can be companies, movements, people or ideas (Mohammad et al., 2016b). It was initially applied to the analysis of political debates in online forums and has become very attractive to measure public opinion on social networks (Aldayel and Magdy, 2019). SD on social media can be categorized based on different criteria, including the type of target, the type of stance (i.e., in favor, against, or neu- tral), and the level of analysis (i.e., post level or network level). The features used for classifica- tion vary according to the analysis level: textual features only (post level) or user-related attributes and behaviors such as mentions and the number of followers (network level) to improve the model accuracy (ALDayel and Magdy, 2021). The state-of-the-art methods for SD (Al-Ghadir et al., 2021; Lai et al., 2017) are based on Machine Learning (ML) and have shown to be effective in various scenarios (Aldayel and Magdy, 2019). However, they rely on manual and complex fea- ture engineering, particularly when applied at the network level. On the other hand, Deep Learning (DL) based methods for SD (Siddiqua et al., 2019; Li and Caragea, 2019) do not require feature engi- neering, but they can easily overfit if not trained with enough labeled data, due to their high number of parameters (Han et al., 2021). Unfortunately, labeling data is an expensive and time-consuming task, leading to small labeled datasets for specific domains (Al-Ghadir et al., 2021). Transfer learning (Zhang et al., 2020; Giorgioni et al., 2020) and unsupervised approaches (Dar- wish et al., 2020; Rashed et al., 2021; Wei et al., 2019) are promising directions for SD, but …

Content truncated to 500 words