MexHat: A Dataset for Hate Speech Detection in Mexican Spanish Videos

What it claims

MexHat contains 1,378 one-minute YouTube clips in Mexican Spanish. Every clip has a three-way label: no negative content (NN), offensive (OF) or hate speech (HS). HS clips also carry one of three sub-group labels: hs1 is ethnicity, nationality or race; hs2 is sexism or anti-LGBTQ+; hs3 is everything else, such as classism. The authors believe it is the only video hate-speech resource for this language.

How it works

Collection. The authors combined three YouTube retrieval routes:

• Searches using 56 terms from the Mexican foreign ministry's inclusive-language guide and 29 Mexico-specific slurs from Hatebase.
• Manually chosen seed videos (20 minutes or shorter, mostly stand-up, opinion videos and soap operas) expanded through the YouTube API.
• Hand-picked channels and playlists.

They cut the videos into 1-minute clips, which gave about 8k clips.

Annotation. Sixteen annotators labelled clips across three campaigns, with two independent annotators per clip. Hate speech follows the UN definition. Annotators dropped 10.51% of the 1,540 labelled clips as irrelevant: shorter than 10 seconds, not Spanish, or only music.

Label resolution. Disagreements are not adjudicated. The more severe label wins: HS beats OF, which beats NN. In practice, a single annotator's HS vote makes a clip HS. The authors chose this to guarantee examples in every class.

Features shipped for baselines.

flowchart LR
  V[1-min clip] --> W[Whisper medium transcript] --> B[BETO 768-d]
  V --> I[I3D RGB+flow 1024+1024]
  V --> G[VGGish audio 128-d]
  B --> L[LSTM + pool to n=50 + FC]
  I --> L
  G --> L
  L --> C[3-way / fine-grained classifier]

Each modality goes through an LSTM, an AdaptiveAvgPool1d and a linear layer, which bring every sequence to 50 steps. Text is projected from 768 to 384 dimensions, and video (RGB and flow concatenated) to 1024. Average sequence lengths are 59 for audio, 177 for text and 25 for video. Some clips have 0 video descriptors.

Evidence

• Annotator agreement (linear κ): 0.416 on the three-class task and 0.514 on the fine-grained task. About 466 clips (about 34%) had conflicting votes, most often NN vs OF and then NN vs HS.
• Class separation: t-SNE plots of the test split show only slight separation between NN, OF and HS, for each modality alone and for all three concatenated.
• Baselines: the authors say text-including configurations performed better. The fetched page ends before the results section, so no baseline numbers are available here.

How to use it

The dataset, its fixed train/val/test partitions and the multimodal features (transcriptions, audio, video) are at https://github.com/iltocl/MexHat.git. The page states no licence or usage terms. The data comes from YouTube clips, so check the repo for what is distributed and under what terms.

Limits

• HS is tiny for the fine-grained task. The test split has 6 hs1 clips, so per-class scores there will be very noisy.
• Labels lean toward harm. The max-severity merge plus κ≈0.42 means HS and OF include clips that one of two annotators judged benign. Re-aggregating requires the raw per-annotator votes, and the page does not say whether those are released.
• Annotator pool is narrow. It was mostly female, STEM-background and aged 25–29. Most annotators did not state LGBTQ+ identity.
• Content sources are skewed. Keyword- and seed-driven retrieval leans toward stand-up, sketches, opinion videos and soap operas.
• Text features carry ASR error. They come from Whisper medium transcripts. The English translations shown in the paper were produced by Gemini 3.5 Flash.
• This is a preprint submitted to CIARP 2026, and the fetched text is truncated before experiments and results.