Guiding Computational Stance Detection with Expanded Stance Triangle Framework

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 3987–4001 July 9-14, 2023 ©2023 Association for Computational Linguistics Guiding Computational Stance Detection with Expanded Stance Triangle Framework Zhengyuan Liu†, Yong Keong Yap‡, Hai Leong Chieu‡, Nancy F. Chen† †Institute for Infocomm Research (I2R), A*STAR, Singapore ‡DSO National Laboratories, Singapore {liu_zhengyuan,nfychen}@i2r.a-star.edu.sg {yyongkeo,chaileon}@dso.org.sg Abstract Stance detection determines whether the author of a piece of text is in favor of, against, or neu- tral towards a specified target, and can be used to gain valuable insights into social media. The ubiquitous indirect referral of targets makes this task challenging, as it requires computational solutions to model semantic features and infer the corresponding implications from a literal statement. Moreover, the limited amount of available training data leads to subpar perfor- mance in out-of-domain and cross-target sce- narios, as data-driven approaches are prone to rely on superficial and domain-specific features. In this work, we decompose the stance detec- tion task from a linguistic perspective, and in- vestigate key components and inference paths in this task. The stance triangle is a generic linguistic framework previously proposed to describe the fundamental ways people express their stance. We further expand it by character- izing the relationship between explicit and im- plicit objects. We then use the framework to ex- tend one single training corpus with additional annotation. Experimental results show that strategically-enriched data can significantly im- prove the performance on out-of-domain and cross-target evaluation. 1 Introduction Stance (and its variant stancetaking) is a concept defined as a linguistically articulated form of so- cial action whose meaning is construed within lan- guage, interaction, and sociocultural value (Biber and Finegan, 1988; Agha, 2003; Du Bois, 2007; Kiesling, 2022). Its subject can be the speaker in a conversation or the author of a social media post, and its object can be in the form of an entity, concept, idea, event, or claim.1 The stance detection task in natural language processing aims to predict the stance of a piece 1There are various definitions of stance and stancetaking in pragmatics and sociolinguistics field. In this article, we follow the generic “stance act” defined by Du Bois (2007). Text: Service was slow, but the people were friendly. Aspect: “Service” Sentiment: Negative Aspect: “people” Sentiment: Positive Text: I believe in SCIENCE. I wear a mask for YOUR PROTECTION. Target: “wear a mask” Stance: Favor � Target: “Dr. Fauci” Stance: Favor � Target: “no mask activity” Stance: Favor � Target: “CD Disk” Stance: Favor � Table 1: Two examples of aspect-level sentiment anal- ysis and target-aware stance detection. The incorrect label prediction is highlighted in red. The target with implicit mention is highlighted in blue. of text toward specified targets. Stance detection is commonly formulated as a classification prob- lem (Küçük and Can, 2020), and is often applied to analyzing online user-generated content such as Twitter and Facebook posts (Mohammad et al., 2016; Li et al., 2021). When given the text and one specified target (i.e., stance object), …

Content truncated to 500 words