turkish-nlp-suite/TurkishHateMap

Turkish Hate Map - A Large Scale and Diverse Hate Speech Dataset for Turkish

<img src="https://raw.githubusercontent.com/turkish-nlp-suite/.github/main/profile/tuhamalogo.png" width="50%" height="50%">

Dataset Summary

Turkish Hate Map (TuHaMa for short) is a big scale Turkish hate speech dataset that includes diverse target groups such as misogyny, political animosity, animal aversion, vegan antipathy, ethnic group hostility, and more. The dataset includes a total of 52K instances with 13 target groups.

The dataset includes 4 labels, offensive, hate, neutral and civilized.

Here is the distribution of target groups:

All text is scraped from Eksisozluk.com in a targeted manner and sampled. The annotations are done by the data company Co-one. For more details please refer to the [research paper]()

Dataset Instances

An instance looks like:

{
"baslik": "soyleyecek-cok-seyi-oldugu-halde-susan-kadin",
"text": "her susuşunda anlatmak istediği şeyi içine atan kadındır, zamanla hissettiği her şeyi tüketir. aynı zamanda çok cookdur kendisi.",
"label": 2
}

Data Split

Benchmarking

This dataset is a part of SentiTurca benchmark, in the benchmark the subset name is hate, named according to the GLUE tasks. Model benchmarking information can be found under SentiTurca HF repo and benchmarking scripts can be found under SentiTurca Github repo.

For this dataset we benchmarked a transformer based model BERTurk and a handful of LLMs. Success of each model is follows:

For a critique of the results, misclassified instances and more please consult to the [research paper]().

Citation

Coming soon!!