HateClipSeg

This repository provides the dataset introduced in our paper: HateClipSeg: A Segment-Level Annotated Dataset for Fine-Grained Hate Video Detection.

To advance research in temporal hate speech detection, we present HateClipSeg — a large-scale multimodal dataset featuring both video-level and segment-level annotations. It contains over 11,714 video segments, each labeled as Normal or one or more of five Offensive categories:

• Hateful
• Insulting
• Sexual
• Violent
• Self-Harm

Additionally, each segment is annotated with explicit target victim labels, enabling nuanced, victim-aware detection tasks.

---

📊 Dataset Statistics

We maintained a balanced distribution between Offensive and Normal segments. Summary:

Label	Video Count	Segment Count
Hateful	194	2,363
Insulting	280	2,920
Sexual	69	372
Violent	192	1,281
Self-Harm	18	39
Offensive*	380	5,223
Normal	55	6,491

* Offensive categories may co-occur; video counts are non-exclusive.

---

✅ Annotation Process and Quality

Our three-stage annotation protocol — Annotation → Discussion → Re-annotation — yields high inter-annotator agreement, especially at the segment level, improving upon prior hate video datasets that lacked detailed agreement reporting.

Annotation Task	Before Discussion	After Discussion
Video-Level Offensive/Normal	0.791	0.817
Segment-Level Offensive/Normal	0.715	0.757
Offensive Category Label	0.840	0.899
Target Victim Label	0.716	0.721

(Krippendorff’s Alpha)

---

🧠 Task Applications

Thanks to high-quality segment-level labels, HateClipSeg supports advancing hate video detection into the temporal domain, enabling:

1. Trimmed Video Classification Predict a single label per pre-segmented clip.

2. Temporal Video Localization Detect labels with precise start and end timestamps in untrimmed videos.

3. Online Video Classification Real-time prediction on streaming video.

---

Dataset File Structure

Dataset/video_level_annotation.csv

Contains video-level annotations with columns:

• Video ID: Combines platform and video identifier. Platforms:

  • bitchute (prefix bit)
  • youtube (prefix yt)

• Video-Level Label: One or more labels from: ["normal", "hateful", "insulting", "sexual", "violence", "harm"] All except "normal" are offensive categories and support multi-labels.

• Target Victim: One or more target groups mentioned, chosen from 21 predefined categories such as "Woman", "Man", "Asian", "Arab", "Latino", "Black", plus "Other". Multiple selections allowed.

Example

Video ID	Video-Level Label	Target Victim
yt_abc123xyz	[hateful, insulting, violence]	[Woman, Black]

• A YouTube video (abc123xyz) labeled as hateful, insulting, and violent, targeting Women and Black individuals.

---

Dataset/segment_level_annotation.csv

Contains segment-level annotations with columns:

• Video ID: Same format as above.

• Segment-Level Labels: List of multi-hot vectors for each segment. Label indices: 0: normal, 1: hateful, 2: insulting, 3: sexual, 4: violence, 5: harm

• Segment Timestamps: List of [start, end] timestamps per segment.

Example

Video ID	Segment-Level Label	Segment Timestamp
yt_abc123xyz	[[0,1,0,0,0,0], [1,0,0,0,0,0], [0,0,1,0,0,0], [0,0,0,0,1,0], [1,0,0,0,0,0]]	[[0,10], [10,20], [20,30], [30,40], [40,50]]

• Segment labels correspond to:

  • 0–10s: hateful
  • 10–20s: normal
  • 20–30s: insulting
  • 30–40s: violence
  • 40–50s: normal

---

lexicons.json

This JSON file contains the hate lexicons used to search and identify relevant videos across platforms.