Selecting features with L1 regularization in Conditional Random Fields Sélection de caractéristiques pour les champs aléatoires conditionnels par pénalisation L1

Discriminative probabilistic models are able to cope with enriched linguistic representations, typically in the form of extremely large feature vectors. Working in high dimensional spaces is however problematic, and these problems are compounded in the case of structured output models, such as conditional random elds (CRF). In this context, feature selection techniques help building more compact and ef cient models. In this work, we propose a novel estimation algorithm for CRF with L1 penalization, which yields sparse representations, thus implicitly selecting relevant features. We also report experiments conducted on two standard language engineering tasks (chunking and named Entity recognition), for which we analyze the generalization performance and the patterns of selected features. We nally suggest various implementation speed-ups that should allow to ef ciently tackle even larger feature vectors.