Infusing Knowledge from Wikipedia to Enhance Stance Detection

Abstract

Stance detection infers a text author's attitude towards a target. This is challenging when the model lacks background knowledge about the target. Here, we show how background knowledge from Wikipedia can help enhance the performance on stance detection. We introduce Wikipedia Stance Detection BERT (WS-BERT) that infuses the knowledge into stance encoding. Extensive results on three benchmark datasets covering social media discussions and online debates indicate that our model significantly outperforms the state-of-the-art methods on target-specific stance detection, cross-target stance detection, and zero/few-shot stance detection. 1

Zero-shot stance detection: Paradigms and challenges

“…For zero-shot MANY-TOPIC stance on VAST, including external knowledge is the most successful technique (Table 2). Interestingly, incorporating knowledge from Wikipedia (He et al, 2022;Zhu et al, 2022) is substantially better than incorporating commonsense knowledge (Zhang et al, 2020;Liu et al, 2021). Models adding external knowledge through task pre-training (Baly et al, 2020;) also perform well, achieving the best performance on all topics, including non-zero-shot ones (i.e., All F1).…”

“…External knowledge is often drawn explicitly from an external source (e.g., Wikipedia articles related to a topic He et al, 2022;Zhu et al, 2022, commonsense knowledge bases Liu et al, 2021, sentiment and emotion lexica Zhang et al, 2020) and then used either as graphs (Zhang et al, 2020;Liu et al, 2021) or as raw-text that is passed with the input to a language model encoder (He et al, 2022;Zhu et al, 2022). Alternatively, knowledge can be incorporated indirectly through task pretraining (e.g., on ideology prediction Baly et al, 2020;.…”

“…CC and A are not used in cross-topic stance detection because they have no related topic for training. Models are: TGA-Net (Allaway and McKeown, 2020), DTCL (Liu R. et al, 2022), JointCL (Liang et al, 2022b), VTN (Wei and Mao, 2019), PT-HCL (Liang et al, 2022a), Cross-Net (Zhu et al, 2022), prior-bin:gold (Jayaram and Allaway, 2021), TPDG (Liang et al, 2021), TOAD (Allaway et al, 2021), POLITICS , Baly (Baly et al, 2020), SEKT (Zhang et al, 2020), CKE-Net (Liu et al, 2021), TarBk-BERT (Zhu et al, 2022), and WS-BERT-Single (He et al, 2022) Frontiers in Artificial Intelligence frontiersin.org language range from ∼250 to ∼1k instances. Similarly, in xstance there are only 1456 instances in the zero-shot language (i.e., Italian), of which only 283 also have zero-shot topics, compared to ∼11k test instances with zero-shot topics in seen languages (German and French).…”

Stance detection for online public opinion awareness: An overview

“…However, as the scope of network public opinion situation awareness is increasingly wide and the topic concerned is also constantly broad, this strong correlation will no longer exist, thus bringing greater challenges to position detection. Content, network interaction, and background knowledge are important clues for stance detection. Existing researches realized it, especially recent researchers have tried to integrate background knowledge by means of knowledge graphs or Wikipedia documents, 14,15,70 but deal with them independently. How to integrate them to improve stance detection is really challenging problem.…”

“…Content, network interaction, and background knowledge are important clues for stance detection. Existing researches realized it, especially recent researchers have tried to integrate background knowledge by means of knowledge graphs or Wikipedia documents, 14,15,70 but deal with them independently. How to integrate them to improve stance detection is really challenging problem.…”