Low-Resource NLP · 2024
Hate-speech detection for Amharic
Custom NLP models that detect hate speech and misinformation in languages off-the-shelf models cannot handle.
- Role
- ML Engineer
- Year
- 2024
- Stack
- TensorFlow, Python, NLP
Content moderation models that work beautifully in English collapse on Amharic and other low-resource Ethiopian languages. The training data barely exists, the scripts and morphology differ sharply from high-resource languages, and commercial APIs simply do not cover them. At AdotSense I built moderation models for these languages from the ground up.
Working from refined social media data, I trained TensorFlow text-classification models to recognize hate speech and misinformation patterns, then deployed them inside an automated content analysis system that ran against live content. The pipeline covered the full lifecycle: data refinement, training, evaluation, and production serving behind a Python API.
Very few ML engineers anywhere have shipped production NLP for low-resource African languages. It remains the work I am proudest of, and it shaped how I think about building AI where the convenient defaults do not exist.
> key_details
- Custom text-classification models where no off-the-shelf coverage exists
- Full lifecycle ownership from data refinement to production serving
- Automated content analysis system running against live social media
- A globally rare specialization in low-resource African language NLP