ry-rousseau's picture
add classes and accuracy metrics
985c57f verified
metadata
license: mit
datasets:
  - copenlu/mm-framing

RoBERTa topic classifier for topic injection into the Longformer Framing Classifier. Classifies input text into one of 19 discrete topics:

  1. Business & Economy
  2. Crime & Safety
  3. Disaster & Accidents
  4. Education
  5. Entertainment
  6. Environment & Nature
  7. Health
  8. Immigration
  9. Infrastructure & Transport
  10. Legal
  11. Lifestyle & Culture
  12. Media
  13. Other/Unknown
  14. Politics
  15. Science & Technology
  16. Social Issues
  17. Sports
  18. War & Conflict
  19. Weather

These were derived empirically by consolidating the unstructured gpt_topic field from the mm_framing silver dataset into
discrete categories based on similarity.

Achieved a 76.4% validation accuracy on 64,000 examples, which was deemed sufficient for assisting domain-specific reasoning in downstream model.