Multimodal Analysis of State-Funded News Coverage of the Israel-Hamas War on YouTube Shorts

Abstract: YouTube Shorts have become central to news consumption on the platform, yet research on how geopolitical events are represented in this format remains limited. To address this gap, we present a multimodal pipeline that combines automatic transcription, aspect-based sentiment analysis (ABSA), and semantic scene classification. The pipeline is first assessed for feasibility and then applied to analyze short-form coverage of the Israel-Hamas war by state-funded outlets. Using over 2,300 conflict-related Shorts and more than 94,000 visual frames, we systematically examine war reporting across major international broadcasters. Our findings reveal that the sentiment expressed in transcripts regarding specific aspects differs across outlets and over time, whereas scene-type classifications reflect visual cues consistent with real-world events. Notably, smaller domain-adapted models outperform large transformers and even LLMs for sentiment analysis, underscoring the value of resource-efficient approaches for humanities research. The pipeline serves as a template for other short-form platforms, such as TikTok and Instagram, and demonstrates how multimodal methods, combined with qualitative interpretation, can characterize sentiment patterns and visual cues in algorithmically driven video environments.

• The paper presents a multimodal pipeline integrating robust speech transcription, aspect-based sentiment analysis, and visual scene classification to assess conflict coverage.
• It demonstrates that task-specific finetuning and domain adaptation, exemplified by DeBERTa-v3-base’s performance, can outperform larger transformer models.
• Empirical findings reveal pronounced sentiment polarization and varied visual framing across outlets, highlighting significant media bias in short-form news.

Multimodal Analysis of State-Funded News Coverage of the Israel-Hamas War on YouTube Shorts

Overview

This paper presents a systematic, multimodal pipeline to analyze how state-funded news outlets cover the Israel-Hamas war in short-form video content, specifically YouTube Shorts. The approach integrates automatic speech transcription, dependency-based aspect linking for sentiment analysis (ABSA), and visual scene-type classification using advanced visual LLMs (VLMs). The study focuses on four major outlets—Al Jazeera, BBC, Deutsche Welle, and TRT World—across a corpus of over 2,300 Shorts and more than 94,000 visual frames, capturing linguistic and visual signals in algorithmic video environments over a one-year period. The pipeline enables quantitative assessment of sentiment toward salient aspect categories and the semantic distribution of visual cues, supporting effective comparative analysis across outlets and over time.

Multimodal Pipeline and Methodological Contributions

The pipeline leverages Whisper large-v3 for robust automatic transcription of spoken content in Shorts, surpassing native YouTube captions in accuracy and reliability under noisy conditions. Textual content undergoes structured parsing via a biaffine dependency parser (SuPar, UD-trained), anchoring aspect terms relevant to the conflict (e.g., Israel, Palestine, Hamas, Zionism, specific political actors) and enabling precise extraction of aspect mentions. The pipeline’s aspect lexicon is manually curated, grouped into ten categories with 56 surface forms, maximally covering relevant geo-political and religious entities.

Aspect-based sentiment analysis is performed with several transformer variants, finetuned on an augmented gold-standard dataset representing transcript-based political discourse. Empirical results indicate that DeBERTa-v3-base outperforms both larger transformer variants and specialized ABSA models (macro-F1 = 81.9 vs. 72.5 for Qwen LLM), highlighting the critical importance of domain adaptation and resource-efficient model selection over sheer parameter scale.

For semantic scene classification, the study adopts the open-source Qwen3-VL (4B) model, avoiding commercial solutions for reproducibility. An iterative process refines a scene taxonomy into seven categories, balancing coverage and minimal overlap: combat/military action, destruction/humanitarian crisis, news media/interview, political/diplomatic event, public protest/demonstration, symbolic/religious ritual, and other/unknown. The taxonomy is operationalized through strict assignment rules for frame-level classification, supporting salient visual patterns in conflict-related coverage.

Evaluation Results

Manual evaluation confirms high pipeline reliability: Whisper-transcribed sentences exhibit accurate segmentation and entity resolution in most cases, with challenges arising mainly in shouted or overlapping speech (e.g., rally chants). DeBERTa-v3-base ABSA displays strong true positive rates for highly politicized aspect terms—even with limited training instances—and maintains robust performance for newly introduced, low-frequency categories. However, the model struggles with ambiguous context or surface-level cues, especially for neutral sentiment in contested statements, reflecting inherent limitations of automated sentiment interpretation.

Qwen3-VL achieves a scene-type classification accuracy of 86.9% (694/799 frames), excelling in discriminating news media/interview settings and formal political events, but showing increased false positives in the residual other/unknown category due to the broad semantic scope and visual ambiguity. Visual overlap in destruction/humanitarian crisis vs. combat/military action is noted, particularly in the aftermath of active conflict.

Empirical Findings

Sentiment Patterns

Cross-outlet analysis of sentiment distributions reveals strongly polarized coverage. Al Jazeera and TRT World consistently favor negative sentiment toward Israel and positive sentiment toward Palestine and related aspects, while BBC and DW maintain predominantly neutral sentiment distributions. Across all outlets, 43.7% of transcripts are non-neutral; negative framings occur more frequently than positive, indicating a general tendency toward negative emotional toning. Notably, sentiment polarity correlates with view counts: positive framing in AJ and TRT yields higher median views, whereas BBC’s negative videos attract more engagement.

Temporal analysis shows persistent negative framing of Israel and positive framing of Palestine throughout the study period, unaffected by evolving real-world events. Explicit polarity is frequently embedded through on-scene signals (e.g., protest chants), quotations from interviewees, or overlay text, rather than anchor narration. Zionism mentions remain stably negative; positive evaluation of Israeli politicians is rare and mostly limited to direct quotations from other political leaders.

Visual Scene Dynamics

A uniform frame sampling strategy captures the visual distribution of semantic scene types. News media/interview scenes dominate (34.7%), with destruction/humanitarian crisis images (16.4%) and political/diplomatic events (11.8%) following. Dramatic spikes in destruction imagery align with major events (e.g., Khan Yunis crisis, polio outbreak) and coincide with high view engagement, often in videos lacking spoken narration. Public protest scenes account for 12.5%, peaking during campus encampments and major global demonstrations. Politically significant events (e.g., UN assemblies, ICJ hearings) and combat/military action are comparatively rare, though military action becomes prominent in October 2024.

Visual taxonomy captures cross-outlet variation: TRT World disproportionately features destruction and protest imagery, reflecting editorial choices that increase emotional salience. Taxonomy simplifies visual diversity, enabling systematic outlet comparison while capturing dynamic trends corresponding to real-world news cycles.

Strong Claims and Contradictory Evidence

The authors report that smaller, domain-adapted models can outperform larger transformers and LLMs in transcript-based ABSA, contrary to prevalent trends favoring parameter scale. Specifically, DeBERTa-v3-base exhibits superior performance compared to both DeBERTa-v3-large and Qwen2.5-7B (QLoRA-finetuned). This underscores the value of task-specific finetuning and resource-efficient modeling for humanities and computational social science applications.

While prior literature suggests overt multimodal alignment of text and visuals in conflict coverage, this study finds that audio, textual, and visual affective cues are interwoven subtly, rarely producing clear bimodal amplification in short-form news. Emotional cues are primarily conveyed through protest and crisis imagery, with textual polarity mostly reproducing on-scene organic signals or interviewee quotations.

Implications for Practical and Theoretical Research

The pipeline offers a scalable, reproducible method for analyzing digital discourse in short-form video environments, supporting researchers in humanities and computational social science domains with limited computational resources. Its modular structure allows adaptation to other platforms (TikTok, Instagram), languages, and event types, promoting extension to large-scale monitoring and comparative studies.

The empirical findings on affective toning and visual framing highlight the crucial role of recommendation-driven environments and short-form formats in shaping audience perception and engagement. The approach provides a foundation for fine-grained analysis of polarization, rhetorical framing, and multimodal signal integration, informing future work in automated bias detection, propaganda annotation, and media influence studies.

Theoretically, the results challenge assumptions regarding the primacy of model size in applied sentiment analysis and underscore the necessity of domain adaptation. The work also nuances understanding of multimodal alignment in political communication, suggesting the need for refined models capturing subtle interplay rather than overt multimodal amplification.

Future Directions and Developments

Expansion to non-English content and other platforms is flagged as evolving priorities, requiring increased per-outlet video bases and expanded visual taxonomy for multi-language/multi-modal integration. Integration of ABSA and scene analysis at the outlet level could support cross-lingual and cross-platform studies of polarization and bias. Further refinement of scene classification prompts and taxonomy, especially for ambiguous categories (other/unknown), coupled with qualitative interpretation, promises improved granularity.

Potential risks include the challenge of distinguishing authentic discourse from AI-generated or manipulated imagery, demanding robust detection and analysis methods in future research.

Conclusion

This paper introduces a multimodal pipeline for rigorous analysis of conflict-related short-form news coverage, demonstrating the utility of domain-adapted models and semantic taxonomies in dissecting both linguistic and visual affective cues. The approach establishes a methodological framework for comparative analysis across outlets and platforms, providing resource-efficient insights into digital discourse dynamics. Results reveal strong polarization and emotional framing in both text and visuals, shaped subtly through multimodal interplay, and highlight the practical value of task-adaptive modeling for real-world applications in the study of online news and political communication.