Hardcoding airpods (and other stories from NLP in insurance)
summary
Worked with real Norwegian claims data and ships concrete solutions: built a “Netflix of unhappy customers” Streamlit app that ranks messages using classifiers trained on a silver dataset from weak supervision labeling functions, improved ranking by replacing cross-entropy with focal cross-entropy for better calibration, summarized clusters using BERTopic plus LLM-generated labels, and solved a 14% automation gap by embedding the object tree for exhaustive search — with one surprising fix being a hardcoded "airpods"→headphone rule that fixed an overrepresented evaluation edge case.
AI-written summary · about
chapters (8)
- 0:00Introduction and background
- 2:10Customer communication insights
- 7:09Weak supervision for labeling
- 12:05Model calibration challenges
- 15:55Topic modeling for insights
- 18:25Automation in claims processing
- 23:00Airpods classification problem
- 26:13Q&A session
related talks
24mChallenges in building Machine Learning models in production
32mData science vs real world: friends or foes
30mDriving a Higher Level of Customer Support with Machine Learning
21mFrom 0 to a crowd sourced neural network in a jiffy
45m"Breaking black-box machine learning"
41mSpotlight on AI use cases and innovators around the world
33mImpact of NLP in AI-powered Bank
35mImproving Credibility of ML Solutions