Extracting Conflict Events from Radio-Based News

Overview

An end-to-end NLP pipeline that monitors live African radio broadcasts (in the DRC, Uganda, and Ethiopia) to automatically extract and structure geopolitical conflict events.

Technologies Used

• Languages: Python, PyTorch, Hugging Face Transformers
• Audio Processing: FFmpeg, Silero VAD, PyAnnote (Speaker Diarization)
• Speech-to-Text (STT): Sunbird AI, Meta MMS-1B-All, Meta MMS-LID-1024
• Translation & NER: Google Translate API, GLiNER (Zero-shot NER)
• LLMs: Llama 3.3 (70B) via Groq API

Pipeline Architecture

Full Technical Report

You can read the detailed project report, which includes the full methodology, experimental results, and comprehensive limitation analysis, here: Read the Full Technical Report.

How to Run (Google Colab)

This pipeline is optimized for Google Colab. To run it:

• Open RadioEventsExtractionPipeline.ipynb in Google Colab
• Add the following keys where needed in the cells:
  • GROQ_API_KEY
  • SUNBIRD_API_TOKEN
  • HF_TOKEN
• Run all cells

Output

For every verified conflict event detected, the pipeline automatically generates three files:

• [country]audio[id].wav: The cleaned, 5-minute raw audio recording
• [country]transcript[id].txt: The full, diarized text transcript before translation
• [country]event[id].txt: The full report for the 5 minute recording, including the stream name, the raw transcript, the translated transcript, the GLiNER entities, and the Structured Events