Robust Integration of Contextual Information for Cross-Target Stance Detection

Proceedings of the The 12th Joint Conference on Lexical and Computational Semantics (*SEM 2023), pages 494–511 July 13-14, 2023 ©2023 Association for Computational Linguistics Robust Integration of Contextual Information for Cross-Target Stance Detection Tilman Beck∗1, Andreas Waldis∗1,2, Iryna Gurevych1 1Ubiquitous Knowledge Processing Lab (UKP Lab) Department of Computer Science and Hessian Center for AI (hessian.AI) Technical University of Darmstadt 2Information Systems Research Lab Department of Computer Science, Lucerne University of Applied Sciences and Arts www.ukp.tu-darmstadt.de www.hslu.ch Abstract Stance detection deals with identifying an au- thor’s stance towards a target. Most existing stance detection models are limited because they do not consider relevant contextual infor- mation which allows for inferring the stance correctly. Complementary context can be found in knowledge bases but integrating the context into pretrained language models is non-trivial due to the graph structure of standard knowl- edge bases. To overcome this, we explore an approach to integrate contextual information as text which allows for integrating contextual in- formation from heterogeneous sources, such as structured knowledge sources and by prompt- ing large language models. Our approach can outperform competitive baselines on a large and diverse stance detection benchmark in a cross-target setup, i.e. for targets unseen during training. We demonstrate that it is more ro- bust to noisy context and can regularize for un- wanted correlations between labels and target- specific vocabulary. Finally, it is independent of the pretrained language model in use.1 1 Introduction Given a text and a target the text is directed at, stance detection (SD) aims to predict whether the text contains a positive or negative stance towards the target or is unrelated. We provide an example in Figure 1. In contrast to formal polls, stance detec- tion (SD) provides a scalable alternative to assess opinions expressed in unstructured texts. However, in contrast to predicting the polarity of a text (i.e., sentiment analysis), SD is more challenging be- cause it requires establishing the relation towards a target which is rarely mentioned in the text (Au- genstein et al., 2016). Further, to infer the correct stance, often the text alone is not sufficient and contextual information ∗* Equal contribution. 1Data and code at https://github.com/UKPLab/ arxiv2022-context-injection-stance needs to be taken into account (Du Bois, 2007). In contrast, most stance classification models are expected to make a correct prediction given the text and target only. This can lead to overly relying on label correlations with target-specific vocabu- lary (Reuver et al., 2021; Thorn Jakobsen et al., 2021). In our example (Figure 1), it is challenging to follow the reasoning of the text if the meaning of school spirit is left unclear. Target: School Uniforms Label: Pro Text: Creates a sense of school spirit. Context: [’school spirit is the enthusiasm and pride felt by the students of a school’, ’a strong sense of school spirit is a positive and uplifting influence on the school and its students’] Figure 1: Example for Stance Detection from the UKP ArgMin dataset (Stab et al., 2018). The context is not part of the …

Content truncated to 500 words