MultiHateClip: A Multilingual Benchmark Dataset for Hateful Video Detection on YouTube and Bilibili

Abstract: Hate speech is a pressing issue in modern society, with significant effects both online and offline. Recent research in hate speech detection has primarily centered on text-based media, largely overlooking multimodal content such as videos. Existing studies on hateful video datasets have predominantly focused on English content within a Western context and have been limited to binary labels (hateful or non-hateful), lacking detailed contextual information. This study presents MultiHateClip1 , an novel multilingual dataset created through hate lexicons and human annotation. It aims to enhance the detection of hateful videos on platforms such as YouTube and Bilibili, including content in both English and Chinese languages. Comprising 2,000 videos annotated for hatefulness, offensiveness, and normalcy, this dataset provides a cross-cultural perspective on gender-based hate speech. Through a detailed examination of human annotation results, we discuss the differences between Chinese and English hateful videos and underscore the importance of different modalities in hateful and offensive video analysis. Evaluations of state-of-the-art video classification models, such as VLM, GPT-4V and Qwen-VL, on MultiHateClip highlight the existing challenges in accurately distinguishing between hateful and offensive content and the urgent need for models that are both multimodally and culturally nuanced. MultiHateClip represents a foundational advance in enhancing hateful video detection by underscoring the necessity of a multimodal and culturally sensitive approach in combating online hate speech.

• The paper introduces MultiHateClip, a multilingual dataset addressing hate speech detection on YouTube and Bilibili by integrating text, audio, and visual modalities.
• The methodology involves annotating 2,000 video clips for hatefulness, offensiveness, and normalcy, with detailed cultural context focusing on gender discrimination.
• Benchmarking reveals that multimodal models, notably GPT-4V, outperform unimodal approaches, underscoring the need for culturally tailored analysis.

MultiHateClip: A Multilingual Benchmark Dataset for Hateful Video Detection on YouTube and Bilibili

Introduction

The study introduces MultiHateClip, a comprehensive multilingual benchmark dataset aimed at enhancing hateful video detection on platforms such as YouTube and Bilibili. Given the growing influence of social media in disseminating hate speech, this dataset addresses the limitations of existing research focused mostly on text-based hate speech and the Western context. MultiHateClip underscores the necessity of exploring multimodal video content, integrating comprehensive annotations and leveraging both English and Chinese cultural contexts to tackle hate speech effectively.

The dataset constitutes 2,000 short video clips, annotated for hatefulness, offensiveness, and normalcy, along with detailed contextual information including targeted victims and contributing modalities. Through this compilation, the study offers a novel perspective in cross-cultural hate speech analysis, specifically focusing on gender-related discrimination in contrasting cultural settings.

Related Work

Contemporary hate speech detection has primarily revolved around text-based analysis, though efforts such as meme detection have recently gained traction in multimodal contexts. Video-based hate speech detection remains notably underexplored, restricted by the limited availability of datasets. Previous works like Das et al.'s English dataset highlight the domain’s nascent stage, emphasizing simple binary classification without detailed contextual analysis.

The MultiHateClip dataset advances this field by offering detailed segment annotations, identifying target victims, and specifying contributing modalities. This approach not only enriches understanding but also addresses the relatively unexplored area of multimodal detection in non-Western languages.

MultiHateClip Dataset

Data Collection and Annotation

The dataset was compiled from YouTube and Bilibili, targeting gender-based hate lexicons and sourcing video clips of up to 60 seconds in length. Utilizing these hate lexicons allowed for the filtering of videos potentially featuring offensive or hateful content.

For annotation, the process involved categorizing each video as hateful, offensive, or normal, identifying segments containing negative content, and pinpointing the target victim group (e.g., Woman, Man, LGBTQ+). Furthermore, annotators were asked to specify whether the hatefulness derived from textual, audio, or visual components.

Data Statistics and Analysis

The dataset reflects stringent moderation policies on the platforms, with a considerable portion categorized as normal despite intended searches for hate-associated content. The victim group analysis signifies a pronounced focus on gender discrimination, especially targeting women. Different modalities conveyed hate speech's multifaceted nature, particularly in Chinese videos where multimodal contributions were significant.

Figure 1: Amplitude of English YouTube videos. Y-axis: Amplitude Indicator, X-axis: Time(sec.).

Figure 2: Zero Crossing Rate of English YouTube videos. Y-axis: Zero Crossing Indicator, X-axis: Time(sec.).

The text analysis using tf-idf scores revealed prevalent hate lexicons in offensive videos, whereas implicit hatefulness was harder to detect through sole textual analysis. Audio and vision analysis highlighted congruent patterns, with offensive videos generally exhibiting higher sound intensities and a distinct visual profile.

Benchmarking Models

Problem Definition and Models

The dataset posed a challenge in distinguishing between hateful, offensive, and normal categories across modalities. Models evaluated include text-based (mBERT, GPT-4, Qwen), audio-based (MFCC, Wav2Vec), vision-based (ViViT, ViT), and multimodal models (VLM, GPT-4V, Qwen-VL).

The multimodal model combining text, audio, and vision features demonstrated enhanced efficacy compared to unimodal approaches, emphasizing the importance of modality integration for comprehensive hate speech detection.

Figure 3: Framework of the multi-modal model. FC: Fully Connected Layer.

Experimentation and Results

The experiments underscored the effectiveness of GPT-4V and the multimodal models for English data, while V1 and M1 excelled in the Chinese dataset. The results highlighted the need for tailored training on non-Western data to overcome cultural bias. Error analysis revealed the benefits of multimodal approaches, as they could recognize instances of implicit hatefulness to a greater extent.

GPT-4V's expansive training data enabled superior content understanding, a strength not paralleled by other models, which struggled with nuanced and implicit hate speech detection.

Conclusion

MultiHateClip advances the field by addressing the previously unmet needs of multimodal hateful video detection in non-Western languages. This dataset and accompanying benchmarks advocate for integrated multimodal approaches, pushing for adaptations in model training to better handle cultural nuances. As such, MultiHateClip stands as a foundational resource poised to guide future research in detecting and understanding hate speech across diverse, multilingual contexts.