This repository contains sensitive content that may be disturbing to some readers.
The videos and data provided for annotation in this project have been sourced from publicly accessible social media platforms in compliance with all applicable laws and regulations.
The content within these videos does not reflect the opinions, beliefs, or viewpoints of the research group or its members.
This dataset is intended solely for research purposes and must be treated with strict confidentiality. No personally identifiable information is included, and all procedures align with relevant legal and ethical standards.
This is the official repository for “DeHate: A Holistic Hateful Video Dataset for Explicit and Implicit Hate Detection” (MM'25).
DeHate is a large-scale multimodal dataset for the detection of explicit and implicit hate in short-form videos. The dataset comprises over 6,689 annotated videos collected from TikTok and BitChute, spanning six social groups, with fine-grained segment-level labels.
This repository contains:
- The cleaned annotation data in
.xlsxformat. - Documentation and statistics on label distributions, hate segment, contributing modality, and target groups.
DeHate includes:
- Videos from two platforms: TikTok (highly regulated) and BitChute (loosely moderated)
- Annotation of both explicit and implicit hate
- Segment-level localization (timestamps for hateful content)
- Modality labels (text, audio, visual)
- Six target groups: Race, Gender, Religion, Sexual Orientation, Disability, Immigration
Label distribution of DeHate.
| Platform | Hate (Explicit) | Hate (Implicit) | Non-Hate |
|---|---|---|---|
| BitChute | 1055 | 835 | 3162 |
| TikTok | 115 | 115 | 1407 |
| Total | 1170 | 950 | 4569 |
Hate Segment Length and Proportion Distributions
The distributions of hate segment lengths (in seconds) and proportions (%) in videos for each platform and hate type (explicit, implicit), with x- and y-axes in log scale and mean values marked by vertical dashed lines.
Normalized Starting Time Distribution of Hate Segments
The density-normalized starting times of hate segments by platform and video length, comparing explicit and implicit types.
Distribution of contributing modalities (Top 3 in bold).
| Modality | BitChute | TikTok |
|---|---|---|
| Text only | 951 | 23 |
| Visual only | 9 | 15 |
| Audio only | 207 | 66 |
| Visual + Audio | 257 | 92 |
| Visual + Text | 26 | 2 |
| Text + Audio | 82 | 3 |
| Visual + Audio + Text | 358 | 29 |
Modality Composition of Explicit and Implicit Hate
The modality composition for explicit and implicit hate in the dataset.
Distribution of hateful videos by single and multiple target groups (Top 3 single targets in bold).
| Target Group(s) | BitChute | TikTok |
|---|---|---|
| Single Target Group | ||
| Race | 573 | 71 |
| Gender | 289 | 16 |
| Religion | 71 | 28 |
| Sexual Orientation | 203 | 19 |
| Disability | 23 | 31 |
| Immigration | 318 | 33 |
| Top 5 Multiple Target Groups | ||
| Race + Immigration | 134 | 14 |
| Race + Religion | 61 | 6 |
| Race + Gender | 59 | 2 |
| Gender + Sexual Orientation | 46 | 4 |
| Race + Sexual Orientation | 38 | – |
| Race + Religion + Immigration | – | 2 |
Proportion of Hateful Videos by Target Group
The proportion of hateful videos by targeted group, separated into explicit and implicit hate.
Sentiment Scores and Top Keywords in Textual Content
The sentiment scores of textual content across explicit hate, implicit hate, and non-hate categories for BitChute and TikTok (left). The top 10 keywords for each category: explicit, implicit, non-Hate (right).
(a) Zero Crossing Rate (ZCR) for Explicit, Implicit, and Non-Hate Content
The Zero Crossing Rate (ZCR) of the audio signal over time for explicit hate, implicit hate, and non-hate content, respectively.
(b) Audio Emotion Heatmap
The distribution of detected emotions in the audio across different content types, with columns representing emotions and rows representing explicit, implicit, and non-hate categories.
Object Detection and Word Cloud for Explicit and Implicit Hate Content
The detected object categories (word clouds) and example video frames with object detection results for explicit hate (left) and implicit hate (right) content.
Below we report the performance of several unimodal and multimodal baselines evaluated on the DeHate dataset.
Binary (Hate vs. Non-Hate) Results
| Model | Accuracy | Macro F1 | F1(H) | Pre(H) | Rec(H) |
|---|---|---|---|---|---|
| Bert | 0.560 | 0.558 | 0.529 | 0.399 | 0.783 |
| RoBERTa | 0.744 | 0.699 | 0.583 | 0.600 | 0.567 |
| OpenAI_TextEmb | 0.758 | 0.708 | 0.586 | 0.636 | 0.544 |
| MFCC | 0.551 | 0.528 | 0.425 | 0.525 | 0.357 |
| VGGish | 0.528 | 0.528 | 0.522 | 0.384 | 0.818 |
| Whisper | 0.679 | 0.595 | 0.410 | 0.489 | 0.353 |
| ViT | 0.649 | 0.587 | 0.427 | 0.414 | 0.441 |
| ViViT | 0.612 | 0.600 | 0.518 | 0.428 | 0.657 |
| Timesformer | 0.661 | 0.611 | 0.471 | 0.463 | 0.478 |
| T1+A2+V1 | 0.620 | 0.610 | 0.545 | 0.438 | 0.720 |
| T2+A1+V2 | 0.689 | 0.630 | 0.481 | 0.456 | 0.509 |
| GPT4-o | 0.523 | 0.521 | 0.493 | 0.371 | 0.733 |
Multiclass (Explicit, Implicit, Non-Hate) Results
| Model | Accuracy | Macro F1 | F1(E) | Pre(E) | Rec(E) | F1(I) | Pre(I) | Rec(I) |
|---|---|---|---|---|---|---|---|---|
| Bert | 0.492 | 0.400 | 0.306 | 0.281 | 0.335 | 0.251 | 0.181 | 0.397 |
| RoBERTa | 0.683 | 0.488 | 0.427 | 0.402 | 0.455 | 0.221 | 0.370 | 0.158 |
| OpenAI_TextEmb | 0.702 | 0.529 | 0.477 | 0.469 | 0.485 | 0.277 | 0.333 | 0.237 |
| MFCC | 0.660 | 0.288 | 0.068 | 0.043 | 0.169 | 0.000 | 0.000 | 0.000 |
| VGGish | 0.507 | 0.408 | 0.293 | 0.228 | 0.409 | 0.269 | 0.236 | 0.312 |
| Whisper | 0.657 | 0.391 | 0.264 | 0.336 | 0.217 | 0.117 | 0.286 | 0.074 |
| ViT | 0.608 | 0.416 | 0.270 | 0.268 | 0.273 | 0.214 | 0.195 | 0.237 |
| ViViT | 0.575 | 0.401 | 0.207 | 0.236 | 0.185 | 0.259 | 0.223 | 0.309 |
| Timesformer | 0.569 | 0.399 | 0.287 | 0.246 | 0.347 | 0.181 | 0.220 | 0.347 |
| T1+A2+V1 | 0.617 | 0.431 | 0.281 | 0.292 | 0.270 | 0.242 | 0.255 | 0.255 |
| T2+A1+V2 | 0.603 | 0.411 | 0.305 | 0.302 | 0.307 | 0.174 | 0.158 | 0.158 |
| GPT4-o | 0.548 | 0.534 | 0.454 | 0.367 | 0.593 | 0.202 | 0.364 | 0.364 |
DeHate.xlsx— Main data file with video-level and segment-level annotations.
If you would like to access the DeHate dataset, please complete the following application form:
DeHate Dataset Application Form
Once your application has been reviewed and accepted, we will send you a download link for the dataset.
If you have any questions, please contact yuchen.zhang@essex.ac.uk.