Atténuation des surdétections d'un correcteur grammatical de qualité commerciale

We describe a study conducted on the proposal of Druide informatique inc., in collaboration with RALI, aiming at developing a system capable of detecting “overdetections”, i.e. designed for filtering detections erroneously flagged by a grammar checker. Various families of classifiers have been trained in a supervised way for 14 types of detections made by a commercial grade French grammar checker. Eight of the 14 classifiers we devised are now part of the latest edition of the grammar checker, embedded in a popular writing assistant. This project was conducted over a six-month period and is an interesting illustration of how a machine learning component can be successfully embedded in a robust, popular natural language application for commercial use. MOTS-CLÉS : correction grammaticale, classification de détections, ingénierie langagière.