Keywords
ontology, LLM, AFA, Information extraction

Technologies
python, owlready2, numpy, pandas, scikit-learn

This project is about further automating troubleshooting problems. The research focuses on combining LLM information extraction with active feature acquisition. In simple words, the system extracts structured information from text records of previously solved problems and then it uses that knowledge to ask users a series of targeted questions to diagnose an unseen problem and suggest a diagnosis if it is confident in it.

The evaluation shows that the system identifies key features from text records with approximately 90% accuracy. Even when working with these automated extractions, the system performs reliably; in a study of simple medical cases with 6 illnesses featured in the diagnoses, it correctly identified the primary diagnosis 73% of the time and included the correct answer in its top two choices 99% of the time.

You can interact with the system yourself at our showcase day. We have a web demo available where you can answer questions and see the system update its findings in real time.