A content moderation model flags policy-violating posts for human review. It runs at 91% precision and 34% recall. Two thirds of violating content is getting through, and the trust-and-safety team is escalating.
The dataset is heavily imbalanced: about 0.4% of posts violate policy. The model is a fine-tuned text classifier trained on 180,000 labelled examples, of which 4,000 are positive. The team's last three attempts to improve it: a larger model, more training epochs, and class weighting: all moved recall by less than 2 points.
You have read the opening. The rest is free with a Bezen account, and no card is needed.
Behind the sign-in