Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 1 min read

Building Content Monitoring System

ai #nlp #africa #opensource I'm building SENAY, a system that monitors public Amharic social media for harmful narratives around GBV, FGM and sexual health misinformation. It will give a civil society organization a

ai #nlp #africa #opensource

I'm building SENAY, a system that monitors public Amharic social media for harmful narratives around GBV, FGM and sexual health misinformation. It will give a civil society organization a weekly summary of what's spreading with a monitoring options.

Chatbot vs. monitor.
My first idea was an Amharic health chatbot: RAG over documents, a Telegram bot, done. But that's mostly wiring up a model someone else trained. I chose the monitor because it means collecting data, defining labels, fine-tuning, and analyzing failures. That's the harder work, and the more useful kind.

Standalone, not integrated.
The partner organization already has digital products. Instead of working inside a codebase I don't know, I'm building a separate system that only feeds them information. In a short project, that saves weeks.

Platform access limits the scope.

  • Telegram: public channels are easy to read, so I'm starting here.
  • Facebook: only public pages are realistic. Private groups are out of reach.
  • Twitter/X: API costs make it impractical, so I'm skipping it and will say so in the write-up.

"Harmful" shouldn't be defined by the developer alone.
A post discussing FGM and a post defending it are very different things. The partner's staff know their communities, so we're writing the annotation guide together, and I'll measure how often annotators agree before training anything.

Open questions.
I don't know yet whether multilingual models like AfroXLMR will handle slang and code-switching in Amharic, how much labeled data I can realistically get, or whether the weekly digest will actually be used. That last one matters most, since a tool nobody uses is just a demo.

Plan:
compare baselines (keyword matching, zero-shot LLM, fine-tuned models), document failure cases, and release the dataset if ethics review allows.

If you've worked on NLP for Amharic or other low-resource languages, I'd like to hear what challenges you ran into.

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.