DimStance: Multilingual Datasets for Dimensional Stance Analysis
Jonas Becker, Liang-Chih Yu, Shamsuddeen Hassan Muhammad, Jan Philip Wahle, Terry Ruas, Idris Abdulmumin, Lung-Hao Lee, Wen-Ni Liu, Tzu-Mi Lin, Zhe-Yu Xu, Ying-Lung Lin, Jin Wang, Maryam Ibrahim Mukhtar, Bela Gipp, Saif M. Mohammed · 2026
Read DimStance: Multilingual Datasets for Dimensional Stance Analysis by Jonas Becker, Liang-Chih Yu, Shamsuddeen Hassan Muhammad, Jan Philip Wahle, Terry Ruas, Idris Abdulmumin, Lung-Hao Lee, Wen-Ni Liu, Tzu-Mi Lin, Zhe-Yu Xu, Ying-Lung Lin, Jin Wang, Maryam Ibrahim Mukhtar, Bela Gipp, Saif M. Mohammed (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-21483.txt.
Introduction
Jonas Becker and a large team build multilingual datasets for a richer kind of stance analysis. Classic stance work sorts a writer's view on a target into broad camps like favor or against. Real attitudes, though, vary in strength and heat, not just in side.
The team models stance along two smooth scales drawn from emotion science. One scale runs from negative to positive, and the other from calm to active. Their datasets cover several languages. Finer scales let researchers track how public feeling shifts in degree, which blunt labels can miss entirely.
Catalog record
| CREATOR | Jonas Becker, Liang-Chih Yu, Shamsuddeen Hassan Muhammad, Jan Philip Wahle, Terry Ruas, Idris Abdulmumin, Lung-Hao Lee, Wen-Ni Liu, Tzu-Mi Lin, Zhe-Yu Xu, Ying-Lung Lin, Jin Wang, Maryam Ibrahim Mukhtar, Bela Gipp, Saif M. Mohammed |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 42 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-21483
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
DimStance: Multilingual Datasets for Dimensional Stance Analysis Jonas Becker, Liang-Chih Yu, Shamsuddeen Hassan Muhammad, Jan Philip Wahle, Terry Ruas, Idris Abdulmumin, Lung-Hao Lee, Wen-Ni Liu, Tzu-Mi Lin, Zhe-Yu Xu, Ying-Lung Lin, Jin Wang, Maryam Ibrahim Mukhtar, Bela Gipp, Saif M. Mohammed Abstract Stance detection is an established task that classifies an author’s attitude toward a specific target into categories such as Favor, Neutral, and Against. Beyond categorical stance labels, we leverage a long-established affective science framework to model stance along real-valued dimensions of valence (negative–positive) and arousal (calm–active). This dimensional approach captures nuanced affective states underlying stance expressions, enabling fine-grained stance analysis. To this end, we introduce DimStance , the first dimensional stance resource with valence–arousal (VA) annotations. This resource comprises 11,746 target aspects in 7,365 texts across five languages (English, German, Chinese, Nigerian Pidgin, and Swahili) and two domains (politics and environmental protection). To facilitate the evaluation of stance VA prediction, we formulate the dimensional stance regression task, analyze cross-lingual VA patterns, and benchmark pretrained and large language models under regression and prompting settings. Results show competitive performance of fine-tuned LLM regressors, persistent challenges in low-resource languages, and limitations of token-based generation. DimStance provides a foundation for multilingual, emotion-aware, stance analysis and benchmarking. rmTeXGyreTermesX \setCJKmainfont bsmi00lp.ttf Offensive Content Warning: This paper contains examples of offensive language for research purposes. 1 Introduction Stance detection aims to automatically identify an author’s stance from text toward a specific target (e.g., a person, organization, or policy) ( Chandra et al. 1981 ; Mohammad et al. 2016 ; Küçük and Can 2020 ; AlDayel and Magdy 2021 ) . The stance is typically classified into predefined labels such as Favor , Against , or Neither (Neutral) . For example, given the sentence “ We must act now to protect the environment .” and the target environment , a stance detection system should label the stance as Favor . In recent years, a wide range of applications and datasets for stance detection have been proposed ( Hardalov et al. 2022 ; Zhang et al. 2024 ; Pangtey et al. 2025 ) , alongside the emergence of specialized task settings such as multimodal Zhou et al. 2025 ; Zhang et al. 2025 and conversational Ding et al. 2025 ; Marreddy et al. 2025 stance detection. Figure 1: Valence-Arousal space. Illustrative examples of multilingual stance expressions. The blue dots are a scatter of valence-arousal scores, showing a U-shaped distribution. Five examples of the DimStance datasets are described as Language_Domain with their corresponding translation printed in green. Underlined text indicates the target aspect. V#A reports the Valence (V) and Arousal (A) annotation for the target aspect. Despite these advances, existing stance detection work still relies on categorical labels for classification. In many settings, stances are rarely purely “for” or “against”. Even expressions sharing the same stance can exhibit substantial variation in affective tone and intensity, motivating the need to identify the affective states underlying stance expressions. By measuring the emotions that drive stance at scale, we can sharpen our understanding of why opinions harden—and, in turn, inform strategies to mitigate the intense polarization seen in debates such as climate change or politics. To address these limitations, we draw insights from long-established affective science frameworks ( Russell 1980 ; Russell 2003 ) , which model evaluative attitudes along dimensions of valence (displeasure–pleasure/negative–positive) and arousal (sluggish–excited/calm–active). As illustrated by the example stance expressions in Figure 1 , English and Nigerian Pidgin would both receive Favor in a categorical scheme, yet the dimensional approach further distinguishes the former as strong, highly aroused support and the latter as mild, low-arousal approval. Likewise, both German and Chinese correspond to Against , although the former reflects emotionally charged opposition while the latter conveys calm disagreement. This enhanced granularity captures the nuances of evaluative attitudes, facilitating a finer-grained stance analysis across languages and domains. Thus, this enriched form of stance detection can be of considerable value in analyzing and understanding public opinion on complex issues such as climate change, vaccine hesitancy, the role of AI in society, and government policies. To our knowledge, no existing stance detection resources adopt a valence-arousal (VA) representation. To fill this gap, we introduce the DimStance , the first dimensional stance resource manually annotated with continuous VA scores. The datasets contain 11,746 target aspects across 7,365 texts and cover five high- and low-resource languages ( Chinese, English, German, Nigerian Pidgin , and Swahili ) across two domains ( politics and environmental protection ). We introduced a new task, termed dimensional stance regression, which requires systems to predict continuous VA scores for each given target aspect within a stance expression. We evaluate a range of pre-trained language models (PLMs) and large language models (LLMs) under regression and prompting settings and analyze cross-lingual VA patterns. We observe that fine-tuned LLM regressors generally outperform prompted LLMs and PLM regressors. Yet, prompting-based LLMs can serve as a cost-efficient alternative. The results further highlight challenges in low-resource languages and limitations of token-based VA prediction. We make our datasets public 1 1 1 https://github.com/DimABSA/DimABSA2026 , providing a pivotal foundation for multilingual dimensional stance analysis to address language-specific nuances and challenges. Our contributions are summarized as follows: • We present the first DimStance datasets, spanning five high- and low-resource languages and two domains. • We introduce a dimensional stance regression task that integrates valence–arousal representations into the stance detection framework. • We provide baseline results across PLM regressors, prompted LLMs, and fine-tuned LLM regressors, facilitating future benchmarking for dimensional stance analysis. All of the data and baselines will be made freely available on the project website. Language Domain Source(s) #Ann. Train Dev Test Total RMSE (V#A) English Environmental Protection EZ-STANCE; Reddit 7 922 / 2059 200 / 339 1020 / 1813 2142 / 4211 1.696#2.195 German Politics Wahl-O-Mat Archive 8 683 / 1335 34 / 75 263 / 438 980 / 1848 1.056#1.818 Chinese Environmental Protection Threads Platform 9 683 / 1091 34 / 49 263 / 898 980 / 2038 0.930#0.902 Nigerian Pidgin Politics X Platform 5 1049 / 1118 119 / 122 331 / 343 1499 / 1583 2.125#2.881 Swahili Politics X Platform 5 1375 / 1622 123 / 145 266 / 299 1764 / 2066 2.958#2.892 Table 1: Dataset overview. For each language, we report the domain, data source(s), total number of annotators, train/dev/test splits, and aggregated totals. All counts are reported as number of texts / number of target aspects . Annotation reliability is measured by root mean squared error (RMSE) for valence#arousal (V#A) ratings. 2 The DimStance Dataset Collection We detail the construction of the dataset, covering data sources, preprocessing steps, and the annotation workflow. 2.1 Data Sources We collect data from five languages, chosen to provide a balance between high-resource languages (English, German, and Chinese) and low-resource languages (Nigerian Pidgin and Swahili). We focus on the domains of politics and environmental protection because they contain a mixture of intense and calm emotional expressions, as well as diverse stances. English. Data for the training split is collected from the environmental protection domain of EZ-STANCE ( Zhao and Caragea 2024 ) . The dev and test splits are obtained from Reddit texts 2 2 2 https://reddit.com/ . Version: 2025-07-01 , using the same