Fouille de règles d'annotation pour la reconnaissance d'entités nommées

Like many NLP tasks, the question of Named Entity Recognition can be adressed either using a symbolic or a data-centered approach. In this paper, we present a hybrid approach which consists in the adaptation of data mining techniques. Our system, mXS, relies on a sequential hierarchical text mining techniques. It implements a data-centered approach to extract symbolic patterns. Besides, mXS relies on an original strategy of recognition which consists in detecting separately the begining and the ending of entities. This strategy is robust on noisy data, especially when speech disfluences or recognition errors occur. Our system has participated to the ETAPE French-speaking evaluation campaign over conversational speech. This paper describes mXS and reports results obtained on reference data (ESTER 2 and ETAPE). MOTS-CLÉS : reconnaissance d’entités nommées, fouille de données, règles d’annotation.