Organizations: Institute for Analytics and Data Science, University of Essex, Colchester, United Kingdom · Department of Engineering, University of Exeter, Exeter, United Kingdom · Department of Computer Science, University of Exeter, Exeter, United Kingdom · School of Electronic Engineering and Computer Science, Queen Mary University of London, London, United Kingdom
Hateful video detection has become increasingly important with the rapid growth of video-centric social media platforms, given the serious risks that hate speech poses to both individual well-being and social cohesion. Compared with text or static multimodal content, hateful video detection remains underexplored and significantly more challenging, as hateful meaning often arises from complex interactions among multimodal cues, including speech, audio, and visual content. Moreover, such signals are often brief, implicit, and temporally dependent, making them difficult to capture using conventional video-level representations. In this work, we propose CLARA, a clip-level multimodal framework for hateful video detection. Instead of treating a video as a single instance, CLARA models it as a sequence of fine-grained clips, enabling more precise capture of temporally localized hateful signals. We introduce a Mixture-of-Experts clip encoder for adaptive multimodal alignment, a local-global segment contrastive objective to jointly model short-term cues and long-range temporal dependencies, and VLM-derived rationales integrated via a gated Transformer to provide high-level semantic guidance. Extensive experiments on three hateful video datasets demonstrate that CLARA consistently outperforms state-of-the-art methods. Further ablation studies and parameter analyses validate the effectiveness of each component.
Figures & tables
Figure 1. An example of hateful video. The hate speech is not fully revealed by single frame or utterance, but emerges through contextual buildup across multiple clips.
Figure 2. Overall framework of CLARA. An input video is first segmented into utterance-aligned clips, from which multimodal features are extracted and fused into clip-level representations via a MoE encoder. Local-global segment contrastive learning is then applied to enhance temporal consistency across clips. Meanwhile, video-level rationales are generated by a VLM and integrated with the clip representations through a rationale-gated Transformer to produce the final video representation, which is finally used for hateful video classification.
HateMM
MHC_CN
MHC_EN
DeHate
Hateful videos
427
298
296
2,120
Non-hateful Videos
638
608
607
4,568
Total videos
1,065
906
903
6,688
Platforms
BitChute
Bilibili
YouTube
BitChute, TikTok
Table 1. Dataset Statistics.
Method
HateMM
MHC_CN
MHC_EN
DeHate
Acc
M-F1
Acc
M-F1
Acc
M-F1
Acc
M-F1
BERT
0.724 ± 0.028
0.709 ± 0.032
0.696 ± 0.006
0.635 ± 0.019
0.674 ± 0.010
0.417 ± 0.020
0.651 ± 0.006
0.522 ± 0.030
Roberta
0.729 ± 0.014
0.718 ± 0.025
0.671 ± 0.003
0.402 ± 0.001
0.672 ± 0.002
0.402 ± 0.001
0.683 ± 0.000
0.406 ± 0.000
MFCC
0.653 ± 0.048
0.598 ± 0.066
0.646 ± 0.014
0.511 ± 0.020
0.656 ± 0.018
0.463 ± 0.031
0.677 ± 0.003
0.441 ± 0.030
Wav2vec
0.696 ± 0.013
0.678 ± 0.016
0.675 ± 0.003
0.403 ± 0.001
0.674 ± 0.008
0.403 ± 0.003
0.683 ± 0.001
0.407 ± 0.001
ViT
0.700 ± 0.031
0.699 ± 0.031
0.701 ± 0.007
0.553 ± 0.024
0.678 ± 0.014
0.438 ± 0.026
0.685 ± 0.010
0.542 ± 0.018
Table 2. Overall Performance in terms of Acc, M-F1, M-Pre and M-Rec. Best in bold and second-best is underlined .
Figure 3. Parameter study of CLARA on HateMM. (a) the effect of the MoE configuration, including the number of experts and top-k routing, (b) the effect of the local and global temporal ratios used in contrastive learning, (c) the effect of the contrastive learning weight λCL , and (d) the effect of the training budget.
Method
HateMM
MHC_CN
MHC_EN
DeHate
Acc
M-F1
Acc
M-F1
Acc
M-F1
Acc
M-F1
w/o A
0.751
0.735
0.704
0.663
0.701
0.667
0.662
0.557
w/o V
0.743
0.739
0.705
0.672
0.688
0.581
0.651
0.510
w/o T
0.712
0.707
0.690
0.653
0.669
0.468
0.668
0.532
w/o MoE
0.789
0.780
0.738
0.679
0.735
0.669
0.708
0.637
w/o CL
0.821
0.802
0.757
0.707
0.749
0.679
0.716
0.641
Table 3. Ablation Results in terms of Accuracy and M-F1.
Figure 4. Expert routing statistics of the MoE. : (a) the distribution measured by gate probability ratio. (b) the distribution of experts measured by top-k selection ratio.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Videos
Total Clips
Clips per Video
Mean
P25
Median
P75
Range
DeHate
6,688
191,554
28.6
6
18
43
1–223
HateMM
1,065
43,564
40.9
8
27
53
1–434
MHC_CN
906
10,158
11.2
3
9
17
1–59
MHC_EN
903
9,676
10.7
4
9
15
1–47
Appendix
Table 4. Distribution of the number of clips per video.
Dataset
Total Clips
Clip Duration (s)
Mean
P25
Median
P75
Range
DeHate
191,554
3.4
1.3
2.2
4.1
0.0 ∗ –254.1
HateMM
43,564
3.4
1.3
2.0
4.0
0.0 ∗ –480.0
MHC_CN
10,158
2.8
1.3
2.0
2.7
0.0 ∗ –58.7
MHC_EN
9,676
3.2
1.5
2.4
4.0
0.0 ∗ –56.4
∗ Values below 0.05 s are rounded to 0.0.
Appendix
Table 5. Distribution of clip durations.
STEP 1 PROMPT (Objective Analysis)
# Role
You are a professional video content verifier. Your task is to accurately summarize what the video shows and what the accompanying text conveys in an objective and neutral way.
# Instructions
You may use:
(1) 20 frames sampled uniformly in temporal order.
(2) Video title (if available) and full transcription.
Appendix
Table 6. Prompt used in Step 1 for objective video analysis.
STEP 2 PROMPT (Hateful Content Verification)
# Role
You are a professional content verifier conducting hateful content analysis. Your task is to make a careful, evidence-based judgment using the information provided.
# Instructions
You may use:
(1) 20 frames sampled uniformly in temporal order.
(2) video title and full transcription.
Appendix
Table 7. Prompt used in Step 2 for hateful content verification.
Element
Description
Content summary
The Content_summary from Step 1.
Visual description
The visual_description from Step 1.
Textual description
The textual_description from Step 1.
Cross-modal relation
A sentence combining cross_modal_relation and cross_modal_explanation from Step 1 in the form “ The relation between textual and visual content is { cross_modal_relation } and { cross_modal_explanation }”. If only one is available, we use it directly.
Contextual elements
If contextually_elements from Step 1 is a non-empty list, we convert it into a sentence of the form “ Contextually elements include: e1,e2,…,ek ”, where ei are the elements in contextually_elements . Otherwise, this field is set to an empty string.
Final decision
A sentence constructed from Label , Explicitness , and Confidence from Step 2. If the Label is non-hate, the sentence “ The video is considered non-hateful with { confidence } confidence. ” If the label is hate, the sentence is “ The video is considered hateful ( { explicitness } ) with { confidence } confidence. ”
Appendix
Table 8. Elements of textual rationale.
Dataset
Metric
CLARA–MoRE
CLARA–HVGuard
CLARA–MM-HSD
HateMM
Acc
4.00e-3
1.63e-2
6.50e-3
M-F1
6.10e-3
2.42e-2
7.50e-3
M-P
2.80e-3
2.42e-2
1.34e-2
M-R
9.40e-3
3.19e-2
1.17e-2
MHC_CN
Acc
6.00e-4
6.50e-3
1.40e-3
M-F1
1.64e-2
3.00e-4
1.24e-2
Appendix
Table 9. Paired t-test p-values comparing CLARA with strong baselines across all datasets.
Figure 5. Parameter study of CLARA on MHC_CN. (a) the effect of the MoE configuration, including the number of experts and top-k routing, (b) the effect of the local and global temporal ratios used in contrastive learning, (c) the effect of the contrastive learning weight λCL , and (d) the effect of the training budget.
Figure 6. Parameter study of CLARA on MHC_EN. (a) the effect of the MoE configuration, including the number of experts and top-k routing, (b) the effect of the local and global temporal ratios used in contrastive learning, (c) the effect of the contrastive learning weight λCL , and (d) the effect of the training budget.
Figure 7. Parameter study of CLARA on DeHate. (a) the effect of the MoE configuration, including the number of experts and top-k routing, (b) the effect of the local and global temporal ratios used in contrastive learning, (c) the effect of the contrastive learning weight λCL , and (d) the effect of the training budget.
Figure 8. Expert routing statistics of the MoE on MHC_CN : (a) the distribution measured by gate probability ratio. (b) the distribution of experts measured by top-k selection ratio.
Figure 9. Expert routing statistics of the MoE on MHC_EN : (a) the distribution measured by gate probability ratio. (b) the distribution of experts measured by top-k selection ratio.
Figure 10. Expert routing statistics of the MoE on DeHate : (a) the distribution measured by gate probability ratio. (b) the distribution of experts measured by top-k selection ratio.
Dataset
Text Encoder
Acc
M-F1
M-P
M-R
HateMM
qwen_0.6
0.864
0.855
0.870
0.846
qwen_8
0.878
0.871
0.882
0.864
bert
0.879
0.872
0.874
0.872
MHC_CN
qwen_0.6
0.780
0.736
0.767
0.730
qwen_8
0.772
0.729
0.764
0.722
bert
0.779
0.734
0.760
0.726
Appendix
Table 10. Impact of text encoder on CLARA across all datasets.
Dataset
VLM
Acc
M-F1
M-P
M-R
HateMM
LLaVA
0.8575
0.8443
0.8577
0.8379
Qwen3VL
0.8790
0.8720
0.8740
0.8720
MHC_CN
LLaVA
0.7084
0.6673
0.6715
0.6514
Qwen3VL
0.7790
0.7340
0.7600
0.7260
MHC_EN
LLaVA
0.7372
0.6898
0.7067
0.6859
Qwen3VL
0.7630
0.7270
0.7320
0.7230
Appendix
Table 11. Impact of rationale generator on CLARA across all datasets.
Dataset
Segmentation
Acc
M-F1
M-P
M-R
HateMM
Fix-length
0.8366
0.8285
0.8345
0.8288
Utterance-aligned
0.8790
0.8720
0.8740
0.8720
MHC_CN
Fix-length
0.7017
0.6547
0.7371
0.6452
Utterance-aligned
0.7790
0.7340
0.7600
0.7260
MHC_EN
Fix-length
0.7497
0.7073
0.7248
0.7041
Utterance-aligned
0.7630
0.7270
0.7320
0.7230
Appendix
Table 12. Impact of clip segmentation strategy.
Method
Params (M)
Training Cost
Inference Cost
Time (s/epoch)
Peak Mem. (MB)
FLOPs (G/video)
Latency (ms/video)
MoRE
110.444
3.961
2,766.45
0.642
1.983
MMHSD
4.626
0.329
352.29
0.008
1.052
HVGuard
3.312
0.205
133.96
0.007
0.552
CLARA
44.950
3.515
2,765.14
2.033
3.720
Appendix
Table 13. Training and inference computational costs on HateMM.
Detecting hateful content in multimodal memes presents unique challenges, as harmful messages often emerge from the complex interplay between benign images and text. We propose GatedCLIP, a Vision-Language model that enhances CLIP's multimodal capabilities with specialized architectural improvements for hateful memes detection. Our approach introduces learned projection heads that map CLIP embeddings to a task-optimized semantic space, a dynamic gated fusion mechanism that adaptively weights visual and textual features, and a contrastive learning objective that maintains cross-modal semantic alignment. Experiments on the Hateful Memes dataset demonstrate that GatedCLIP achieves an AUROC of 0.66, substantially outperforming the CLIP baseline (AUROC 0.49) while maintaining computational efficiency with only 350K trainable parameters.
Hateful videos have become prevalent on online platforms, highlighting an urgent need for effective detection. However, existing studies primarily focus on binary classification and fail to provide contextual rationales that reveal the implicit meanings behind these judgments, significantly undermining model explainability. To fill this gap, we aim to achieve explainable hateful video detection, enabling models to provide contextual rationales that integrate relevant evidence and logical reasoning alongside decisions. This approach can comprehensively enhance the understanding of video content and the explainability of the decision-making process. We first introduce two datasets, Ex-HateMM and Ex-ImpliHateVid, for explainable hateful video detection. Each dataset provides fine-grained annotations of multimodal harmful elements, along with contextual rationales. We then propose an Information Augmentation and Reasoning Enhancement (IARE) framework designed for explainable detection. The framework employs an information augmentation phase that leverages the multimodal chain-of-thought to integrate harmful elements, thereby enriching rationale evidence. Additionally, IARE incorporates a reasoning enhancement phase, in which Direct Preference Optimization guides the model toward correct reasoning paths and away from incorrect ones, thereby improving the logical coherence of its justifications. We conduct extensive experiments on the two datasets, comparing multiple baselines with our proposed IARE framework. The results demonstrate that IARE achieves state-of-the-art performance while also generating accurate rationales.
Junyu Lu, Deyi Ji, Liqun Liu +9
Dalian University of Technology Dalian, China · Tencent Beijing, China · Singapore University of Technology and Design Singapore, Singapore
Combating hate speech on social media is critical for securing cyberspace, yet relies heavily on the efficacy of automated detection systems. As content formats evolve, hate speech is transitioning from solely plain text to complex multimodal expressions, making implicit attacks harder to spot. Current systems, however, often falter on these subtle cases, as they struggle with multimodal content where the emergent meaning transcends the aggregation of individual modalities. To bridge this gap, we move beyond binary classification to characterize semantic intent shifts where modalities interact to construct implicit hate from benign cues or neutralize toxicity through semantic inversion. Guided by this fine-grained formulation, we curate the Hate via Vision-Language Interplay (H-VLI) benchmark where the true intent hinges on the intricate interplay of modalities rather than overt visual or textual slurs. To effectively decipher these complex cues, we further propose the Asymmetric Reasoning via Courtroom Agent DEbate (ARCADE) framework. By simulating a judicial process where agents actively argue for accusation and defense, ARCADE forces the model to scrutinize deep semantic cues before reaching a verdict. Extensive experiments demonstrate that ARCADE significantly outperforms state-of-the-art baselines on H-VLI, particularly for challenging implicit cases, while maintaining competitive performance on established benchmarks. Our code and data are available at: https://github.com/Sayur1n/H-VLI