[论文解读] DimStance: Multilingual Datasets for Dimensional Stance Analysis

Jonas Becker, Liang-Chih Yu+15 more|arXiv (Cornell University)|Jan 29, 2026

Sentiment Analysis and Opinion Mining被引用 0

一句话总结

DimStance 提供首个多语言的情感-唤起注释用于立场分析，引入一个维度立场回归任务，并在五种语言与两个领域上对预训练语言模型和大语言模型进行基准评估。该研究分析跨语言的 VA 模式，突出低资源语言的性能差距以及基于标记的 VA 预测的局限性。

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.

研究动机与目标

• Introduce DimStance, the first dimensional stance resource with valence–arousal annotations across five languages and two domains.
• Provide a dataset of 11,746 target aspects in 7,365 texts to enable dimensional stance regression.
• Define and evaluate a dimensional stance regression task using VA scores.
• Benchmark pretrained language models and large language models under regression and prompting settings to establish baselines.

提出的方法

• Curate DimStance datasets with VA annotations for English, German, Chinese, Nigerian Pidgin, and Swahili in politics and environmental protection.
• Annotate VA scores on target aspects via five native annotators per language and use majority voting for validity.
• Train and evaluate PLM regressors (XLM-R, RemBERT, LaBSE) with a regression head on sentence representations.
• Evaluate closed and open LLMs under prompting (few-shot) and fine-tuned regression setups (LoRA-based with 4-bit quantization).
• Use RMSE on VA (valence-arousal) as the evaluation metric and compare cross-language performance and model families.

(a) English (env. protection)

实验结果

研究问题

• RQ1 How can stance be modeled along continuous valence-arousal dimensions across multiple languages and domains?
• RQ2 What are the cross-lingual patterns in valence-arousal for stance expressions, and how do models perform on dimensional stance regression across languages?
• RQ3 Do fine-tuned LLM regressors outperform prompting-based LLMs and PLMs for dimensional stance prediction, and under what conditions?
• RQ4 What are the challenges and limitations of applying dimensional stance analysis to low-resource languages?
• RQ5 How do token-based VA predictions compare to regression-based VA predictions in terms of distribution alignment and accuracy?

主要发现

• DimStance is the first dataset with manual VA annotations for stance, covering five languages and two domains with 7,365 texts and 11,746 target aspects.
• Fine-tuned LLM regressors generally outperform prompted LLMs and PLM regressors on average, especially for larger 70B models.
• Prompting-based LLMs offer data-efficient baselines but show grid-like, discretized VA outputs that can misalign with continuous VA distributions.
• Cross-language VA patterns reveal language- and domain-specific affective profiles, with Chinese showing compact VA distribution and English/German showing more variation.
• Low-resource languages (Swahili, Nigerian Pidgin) exhibit stronger challenges and larger RMSE gaps, highlighting data scarcity effects.
• Token-based VA prediction in prompting approaches tends to underperform continuous regression-based methods, particularly when VA distributions are compact.

(b) German (politics)