DimStance: Multilingual Datasets for Dimensional Stance Analysis

TL;DR

This work addresses the need for fine-grained stance analysis by introducing DimStance, the first multilingual resource annotated with continuous valence and arousal scores for stance expressions across five languages and two domains. It defines a dimensional stance regression task and benchmarks a spectrum of models including PLMs, prompting-based LLMs, and fine-tuned LLM regressors, highlighting when each approach excels and where challenges persist, especially in low-resource languages. The results demonstrate that VA representations capture nuanced attitudinal states beyond traditional polarity labels and that fine-tuned 70B-scale LLMs offer strong performance, though token-based VA generation and cross-language variability remain hurdles. DimStance thus provides a valuable, publicly available benchmark for multilingual, emotion-aware stance analysis with practical implications for understanding public opinion and polarization while guiding future methodological improvements.

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.

Figures (6)

• 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.
• Figure 2: VA distributions by language. For each language, we show a kernel density plot of the VA scatter space. We show the top ten most prevalent targets in the dataset. An extensive list of the top targets is included in \ref{['tab:top10aspects']}.
• Figure 3: English VA prediction scatter plots. Comparison of gold and predicted VA distributions from selected models on the English test set. The x-axis shows valence, and the y-axis shows arousal on a one to nine scale.
• Figure 4: Trendlines of few-shot performance. We compare the prompting of three closed LLMs. The x-axis represents the number of shots, ranging from zero to a 32-shot setup. The y-axis shows the RMSE.
• Figure 5: Annotation interfaces used in our study.