Adaptive Filtering of Electronic Mail

This paper describes a system for the adaptive filtering of electronic mail. The goal is to detect important messages which are flagged for the user to read in priority. The system uses the Latent Semantic Indexing (LSI) technique to extract “semantic concepts” from a set of sample messages. Adaptation takes place by adding new messages with their evaluation by the user, and eventually modifying the “concept space”. Experiments show that the system is able to detect 10% more important messages than a random classifier. The fixed concept space version has been integrated into a working prototype. This work is part of the MISTRAL project, in collaboration with Telis and Ingenia, with a sponsorship from the French Ministry of Industry.