PaperlayerView exact revision on arXiv

Source-linked AI summary

Visual Framing for News Stance Detection via Image Generation

Dahyun Lee, Jiyoung Han, Kunwoo Park

arXiv:2609.00685v1cs.CLcs.AIcs.CVcs.CY

TL;DR

Article-level news stance detection is challenging because stance cues are implicit and dispersed across long, complex news articles. VFSTANCE uses visual framing to convert those cues into generated images, which are combined with the article for stance prediction. The method outperforms existing approaches, and a controlled study with 200 participants finds that its images improve human stance identification.

  • Problem

    Article-level news stance detection lacks effective ways to identify implicit, subtly framed cues distributed across lengthy, structurally complex news articles.

  • Method

    VFSTANCE uses an LLM to derive visual framing specifications, a T2I model to generate a stance-aware image, and an LVLM to predict stance from the article and image.

  • Results

    VFSTANCE outperforms existing stance detection methods, with visual framing and image generation contributing to performance; its generated images also improve human stance identification in a study with N = 200.

  • Takeaways & Limitations

    Visual framing and image generation can make otherwise implicit stance cues more discernible to computational models and human readers.

  • Takeaways & Limitations

    Multilingual evaluation is limited because the primary testbed is a Korean corpus, supplemented by a German text-only dataset and translated extensions.

Abstract

from arXiv · show

Article-level news stance detection aims to identify the perspective of news articles toward social issues. Despite advances in stance detection and its importance for trustworthy media environments, news articles pose distinct challenges because their stances are often implicit, subtly conveyed through journalistic framing, and embedded in long, structurally complex texts. To address these challenges, we introduce VFStance, which leverages visual framing to make implicit stance cues more explicit via image generation. In evaluation experiments, we demonstrate the effectiveness of VFStance over existing methods and the contribution of visual framing to its performance. Finally, a controlled user study (N=200) in a snippet-based news consumption setting further demonstrates that VFStance can make stance signals visually salient and highlights its potential use beyond automated stance detection.

1 Introduction

Article-level news stance detection is difficult because stances are implicit, dispersed across lengthy complex articles, and not readily captured by methods designed for short texts. VFSTANCE addresses this challenge by using visual framing and image generation to make stance cues more explicit for models and readers.

  • Professional journalistic norms often make article stances implicit and subtly expressed.
  • Long, structurally complex articles distribute framing cues across rhetorically elaborate text, challenging both automated methods and human readers.
  • VFSTANCE uses an LLM to derive visual framing specifications, a T2I model to generate an image, and an LVLM to predict stance jointly from article and image.
  • VFSTANCE outperforms existing approaches on Korean and German article-level stance datasets, while ablations show contributions from visual framing and image generation.
  • N = 200 participants identified article stance more accurately with VFSTANCE-generated images than with alternative image conditions.

2 Related Work

News stance detection has focused largely on short-form or headline-level tasks, while article-level stance toward social issues remains relatively underexplored. Related work also includes multimodal stance detection and computational approaches to visual framing analysis.

  • Most stance detection research examines short-form social media text using pretrained models, LLM reasoning, or in-context learning.
  • Multimodal stance detection methods jointly use text and images, while visual framing analysis studies how visual elements convey interpretive emphasis.
  • News stance detection has largely targeted headlines or sentences in fake-news and rumor-verification settings.
  • Article-level stance toward social issues remains relatively scarce in existing research.
  • Computational framing analysis has become an active research area alongside surveys and manual content analysis using predefined frame categories.

3 Problem and Dataset

The task is to classify an article’s stance toward an issue as supportive, neutral, or oppositional. The study evaluates VFSTANCE on Korean multimodal and German text-only datasets, with additional translated multilingual extensions.

  • Article-level stance detection assigns an article’s stance toward a covered issue as supportive, neutral, or oppositional.
  • K-News-Stance-MM contains 1,816 Korean news articles with accompanying images and provides the study’s primary multimodal testbed.
  • CheeSE contains 1,762 German news articles and supports evaluation in a text-only setting.
  • The study also provides LLM-translated K-News-Stance-MM extensions in English, Chinese, Indonesian, and Arabic.

4 Proposed Method

VFSTANCE is a multi-stage, modular framework that uses visual framing to transform implicit textual stance cues into visually salient signals for article-level detection. It structures framing information, generates a stance-aware image, and combines that image with article text for multimodal prediction.

  • VFSTANCE uses visual framing to convert implicit textual stance cues into explicit signals for multimodal article-level stance detection.
  • Stage 1: Visual Framing Annotation: Stage 1 uses an LLM to produce a structured visual framing specification covering image content and visual presentation.The specification includes actors, objects, scenes, style, composition, angle, distance, saturation, and luminosity.
  • Stage 1: Visual Framing Annotation: The specification organizes ten features across four levels: ideological, connotative, stylistic-semiotic, and denotative.The schema captures both interpretive associations and visually renderable properties.
  • Stage 2: Stance-Aware Image Generation: Stage 2 generates a news image from eight visually renderable features drawn from the stylistic-semiotic and denotative levels.The ideological and connotative levels are excluded from image generation because they represent more abstract interpretive information.
  • Stage 3: Multimodal Stance Detection: Stage 3 gives an LVLM both the article text and generated image to predict stance, while VFSTANCE (TEXT) supplies the framing specification instead of an image.The text variant skips image generation and passes the visual framing specification directly to the detector.

5 Evaluation Results

VFSTANCE outperforms existing article-level stance detection methods, with visual framing and image generation each contributing to performance. Results across ablations and the German CheeSE dataset support the framework’s effectiveness across design choices and languages.

  • VFSTANCE with Gemini-3-flash achieved 0.746 accuracy and 0.747 macro F1, outperforming all evaluated baselines (p<0.01).
  • Ablation: Visual Framing in Image Generation: All three alternative image-generation methods achieved lower accuracy and macro F1 than VFSTANCE, with gaps of at least 0.025.
  • Ablation: Use of Image Generation: VFSTANCE significantly outperformed both visual-framing alternatives without image generation (p<0.01), indicating an additional gain from image generation.
  • Ablation: Visual Framing Level Selection: Denotative and stylistic-semiotic framing levels were critical in Stage 2, while connotative and ideological levels were harder to render into explicit stance signals.
  • Ablation: Visual Framing Level Selection: Annotating all four framing levels in Stage 1 increased macro F1 by 0.01, improving annotations for the denotative and stylistic-semiotic features used in generation.
  • Effectiveness Across Languages: On the German CheeSE dataset, Gemini-3-flash achieved 0.618 accuracy and 0.62 macro F1, outperforming all baseline methods by a substantial margin (p<0.01).The results, together with Korean-language findings, support effectiveness across languages.

6 User Study

A controlled study with 200 native Korean speakers tested stance identification from abbreviated news snippets under four presentation conditions. VFSTANCE-generated images produced the highest accuracy and significantly higher odds of correct identification than text-only, original-image, and naïve-image conditions.

  • Study Design: The study recruited 200 native Korean speakers to classify article stance from snippets containing identical text under different visual presentation conditions.
  • Study Design: Each participant viewed nine articles in randomized order, yielding approximately 450 observations per presentation condition overall.
  • Results: VFSTANCE achieved the highest stance-identification accuracy at 0.378, exceeding text-only, original, and naïve conditions by 0.071, 0.098, and 0.076, respectively.
  • Results: Compared with VFSTANCE, text-only, original, and naïve conditions had lower correct-identification odds of OR = 0.708, 0.620, and 0.697, respectively.All three comparisons were statistically significant: p = 0.001, p < 0.0001, and p = 0.0006, respectively.
  • Interpretation: Despite modest overall accuracy, VFSTANCE-generated images facilitated stance identification under abbreviated news exposure and showed potential beyond automated stance detection.

7 Conclusion

VFSTANCE applies visual framing to transform implicit, dispersed stance cues in long news articles into generated visual representations for stance detection. The framework outperforms existing methods, and its generated images also improve human stance identification in snippet-based consumption.

  • VFSTANCE uses an LLM-produced visual framing specification, a text-to-image model, and an LVLM in a multi-stage stance detection framework.
  • Evaluation results show that both visual framing and image generation contribute to VFSTANCE’s performance.
  • In a controlled snippet-based user study, VFSTANCE-generated images improved stance identification, indicating potential beyond automated stance detection.
  • Visual framing can function as an intermediate representational layer that makes dispersed textual stance cues more accessible to computational models and human readers.
  • The findings point to potential applications of generative AI grounded in visual framing for making media perspectives more transparent and supporting news-bias identification.

Limitations

The study identifies limitations involving computational cost, multilingual coverage, and model selection. These constraints motivate different deployment choices and future evaluation across datasets and model families.

  • Computational Costs: VFSTANCE uses three models across its corresponding stages, creating a computational cost–accuracy trade-off.VFSTANCE (TEXT) omits image generation and remains more efficient, while full VFSTANCE is strongest when maximizing detection accuracy.
  • Multilingual Evaluation: The primary K-News-Stance-MM testbed is Korean, limiting the study’s multilingual evaluation scope.The authors also evaluate German data and translated Korean data in four languages.
  • Model Selection: The primary VFSTANCE backbones are proprietary models selected for stronger language understanding, reasoning, and generation capabilities.Supplementary open-weight experiments achieved lower performance, while the modular framework permits broader future model evaluation.

Ethical Considerations

The study addresses research governance, data access, participant privacy, and risks from synthetic images. It uses gated dataset access, anonymization of private individuals, informed consent, and safeguards for reader-facing image use.

  • Research Governance: The study was approved by the Institutional Review Board at Soongsil University.The approval identifier is SSU-202604-HR-804-1.
  • Copyright and Privacy: K-News-Stance-MM uses gated access and a custom agreement restricting use to non-commercial academic research and prohibiting redistribution.Private individuals are anonymized, while public-figure names remain in the news data.
  • User Study: The user study recruited 200 Korean native speakers, obtained informed consent, and collected no sensitive personal information.Participants could withdraw at any time, and the task took approximately 10 minutes with compensation of about USD 3.6.
  • Risks Associated with Image Generation: Generated images may reproduce or amplify social biases when repurposed for reader-facing applications.The authors recommend labeling images as synthetic and reviewing them for defamation, personality-rights, and synthetic-media regulations.
  • AI Assistant Use: AI-assisted language tools were used exclusively for grammar checking and readability improvements.

A Experimental Setups

The experiments average results over five runs and use specified hardware, software, model configurations, optimization settings, and dataset-split reporting. Appendix materials document model identities, parameter sizes, and dataset distributions.

  • Experimental Protocol: Experimental results are averaged over five runs with standard errors across trainable and API-based models.Trainable models use random seeds 42–46, while API-based models repeat the same inference configuration five times.
  • Experimental Protocol: Experiments use three NVIDIA RTX A6000 GPUs with 48GB each, 128GB RAM, Python 3.10, PyTorch 2.4.1, Transformers 4.57.1, and CUDA 12.1.
  • Baseline Configuration: Visual baselines include ResNet-50, ViT-B/16, and SwinV2-Base, with batch size 32 and early stopping patience 3.Learning rates are 1×10^-4 for ResNet-50 and 5×10^-5 for ViT-B/16 and SwinV2-Base.
  • Baseline Configuration: Multimodal baselines pair KLUE-RoBERTa-large or Korean CLIP with visual encoders and use learning rate 2 × 10^-5 and batch size 16.
  • Image Generation: EAIG4SD follows its original Stable Diffusion 3 Medium setup while replacing unsupported components with Korean-capable alternatives.The configuration uses 28 denoising steps, guidance scale 7, Qwen2.5-VL-7B-Instruct, Korean CLIP, and PageRank selection.
  • Open-Weight Experiments: Open-weight experiments evaluate InternVL3-14B-Instruct and Gemma3-12B-Instruct zero-shot with shared prompts, while Stable Diffusion 3.5 Large generates one image per article.The image generator uses 28 denoising steps and guidance scale 4.5.

B Dataset Details

The study constructs and evaluates article-level stance datasets, compares textual, visual, multimodal, and image-generation methods, and analyzes visual-framing features, costs, modalities, multilingual performance, and ablations.

  • Dataset Details: K-News-Stance-MM contains 1,816 Korean articles with publisher images, preserving an issue-level train/test split to prevent leakage.The dataset has 909 training and 907 testing samples, with supportive, neutral, or oppositional stance labels.
  • Dataset Details: CheeSE provides 1,762 German news articles after excluding unclear and unrelated samples, split into 800/200/762 train, validation, and test examples.It evaluates the method without original news images.
  • Methods and Baselines: Baseline comparisons include textual, visual, multimodal, and image-generation methods using article text, publisher images, or generated images.EAIG4SD generates stance-aware candidate images and selects one using text–image similarity, target consistency, and stance consistency.
  • VFSTANCE: The visual-framing schema covers ideological, connotative, stylistic-semiotic, and denotative levels, including content selection and visual presentation.The stylistic-semiotic level includes style, composition, angle, distance, saturation, and luminosity; connotative and ideological features provide interpretive context but are not rendered in Stage 2.
  • Supplementary Analyses: All models perform substantially worse with images alone than with text, while combining text and images produces the best performance.For GPT-5.4-mini, image-only accuracy is 0.383 and macro F1 is 0.354.
  • Multilingual Evaluation: VFSTANCE achieves higher accuracy and macro F1 than the text-only setting across translated English, Chinese, Indonesian, and Arabic versions.The original Korean version achieves the best overall performance among the evaluated language versions.

D.5 Error Case Analysis

Error analysis shows that VFSTANCE’s changed predictions are concentrated in gold-neutral cases, with supportive shifts linked to generated visual tone and lexical shortcuts. Supplementary open-weight experiments further show that visual framing remains effective across alternative backbones, while Stage 3 is the most sensitive component to model substitution.

  • Prediction changes: 51 of 55 cases where the textual LVLM baseline was correct but VFSTANCE was wrong had gold-neutral labels, with most errors shifting toward supportive.A representative neutral article about reintroducing conscripted police was rendered as a bright patrol scene, shifting the prediction from neutral to supportive.
  • Prediction changes: 48 of 57 cases where VFSTANCE was correct but VFSTANCE (TEXT) was wrong had gold-neutral labels, again with errors skewed toward supportive.A specification containing “bright” caused a supportive misclassification as text, whereas rendering the same specification as a photographic image yielded a correct prediction.
  • Open-weight evaluation: VFSTANCE outperformed corresponding textual, visual, and multimodal baselines under both open-weight LVLM backbones.The supplementary experiments used InternVL3-14B-Instruct, Gemma3-12B-Instruct, and Stable Diffusion 3.5 Large.
  • Stage-level ablation: Replacing the Stage 3 LVLM produced the largest performance drop: 0.127 in accuracy and 0.125 in macro F1.Substitution reduced performance at all three stages, and the fully open-weight variant performed comparably to the Stage 3-only substitution.
  • Pipeline interpretation: The VFSTANCE pipeline converts Stage 1 visual-framing specifications into images through a fixed-template text-to-image prompt before multimodal stance detection.The Stage 1 specifications describe ideological, connotative, and stylistic framing, while Stage 2 fills template placeholders for image generation.
Loading 2609.00685v1…