Using a multilingual dataset to enhance hateful video detection on YouTube and Bilibili

Social media has revolutionized the way information is shared throughout communities, but it can also be a cesspit of hateful content. Current research on hateful content detection focuses on text-based analysis, while hateful video detection remains underexplored.

“Hate speech in videos can be conveyed through body language, tone, and imagery, which traditional text analysis misses. As platforms like YouTube and TikTok reach large audiences, hateful content in video form can be more persuasive and emotionally engaging, increasing the risk of influencing or radicalizing viewers,” explained Roy Lee, Assistant Professor at the Singapore University of Technology and Design (SUTD).

In his paper “MultiHateClip: A multilingual benchmark dataset for hateful video detection on YouTube and Bilibili,” Asst Prof Lee led a team to develop MultiHateClip, a novel multilingual dataset that aims to enhance hateful video detection on social media platforms. He previously created SGHateCheck, a novel functional test evaluating hate speech in multilingual environments. The study is published on the arXiv preprint server.

Using hate lexicons and human annotations focused on gender-based hate, MultiHateClip classifies videos into three categories: hateful, offensive, and normal. Hateful content involves discrimination against a specific group of people based on specific attributes such as sexual orientation.

Offensive content is distressing, but lacks the targeted harm of hate speech and does not incite hatred. Normal content is neither hateful nor offensive. Compared to a binary classification (hateful versus non-hateful), this three-category system allows for a more nuanced approach to content moderation.

After reviewing over 10,000 videos, the team curated 1,000 annotated short clips each from YouTube and Bilibili to represent the English and Chinese languages, respectively, for MultiHateClip. Among these clips, a consistent pattern of gender-based hate against women emerged. Most of these videos used a combination of text, visual, and auditory elements to convey hate, underscoring the need for a multimodal approach to understanding hate speech.

Compared to existing datasets that are simpler and lack detail, MultiHateClip is enriched with fine-grained, comprehensive annotations. It distinguishes between hateful and offensive videos, and outlines which segments of the video are hateful, who the targeted victims are, and what modalities portray hatefulness (i.e., text, visual, auditory). It also provides a strong cross-cultural perspective as it includes videos from both Western (YouTube) and Chinese (Bilibili) contexts, highlighting how hate is expressed differently across cultures.

The team expected distinguishing hateful videos from offensive ones to be difficult as both share similarities, such as inflammatory language and controversial topics. Hateful speech targets specific groups, while offensive content causes discomfort without intending to discriminate. The subtle differences in tone, context, and intent make it challenging for human annotators and machine learning models to draw the line between hateful and offensive content.

“Additionally, cultural and linguistic nuances further complicate the distinction, particularly in multilingual contexts like English and Chinese, where expressions of hate or offense may vary significantly. This complexity underscores the need for more sophisticated detection models that can capture subtle distinctions,” emphasized Asst Prof Lee.

The study also tested state-of-the-art hate video detection models with MultiHateClip. Results highlighted three critical limitations in current models: the difficulty in distinguishing between hateful and offensive content, the limitations of pre-trained models in handling non-Western cultural data, and the insufficient understanding of implicit hate. These gaps emphasize the need for culturally sensitive and multimodal approaches to hate speech detection.

MultiHateClip reflects the value of intersecting design, artificial intelligence, and technology. Its real-world significance is clear—to detect hate speech and prevent its dissemination. Optimized for video content, the model has a cross-cultural focus and is especially useful on social media platforms where videos are the primary form of communication, such as YouTube, TikTok, and Bilibili. Content moderators, policymakers, and educational organizations will benefit from using MultiHateClip to understand and mitigate the spread of hate speech.

“Overall, MultiHateClip plays a crucial role in creating safer, more inclusive online environments,” said Asst Prof Lee, who shared the possibility of collaborating with social media platforms to deploy the model in real-world settings. In addition, the team could potentially look into broadening the dataset to include more languages and cultural contexts, improving model performance by creating better algorithms that can distinguish between hateful and offensive content, and developing real-time hate speech detection tools.