Conférence
Notice
Langue :
Français
Conditions d'utilisation
Droit commun de la propriété intellectuelle
DOI : 10.60527/b3d3-kc18
Citer cette ressource :
CEMU. (2015, 23 juin). 05a - Apprentissage par imitation pour l’étiquetage de séquences : vers une formalisation des méthodes d’étiquetage easy-first (taln2015) , in Session extraction d'information (taln 2015). [Vidéo]. Canal-U. https://doi.org/10.60527/b3d3-kc18. (Consultée le 4 novembre 2024)

05a - Apprentissage par imitation pour l’étiquetage de séquences : vers une formalisation des méthodes d’étiquetage easy-first (taln2015)

Réalisation : 23 juin 2015 - Mise en ligne : 17 juillet 2015
  • document 1 document 2 document 3
  • niveau 1 niveau 2 niveau 3
Descriptif

Sessions orales TALN 2015 – Mardi 23 juin 2015

Session Extraction d’information

Apprentissage par imitation pour l’étiquetage de séquences : vers une formalisation des méthodes d’étiquetage easy-first

Elena Knyazeva, Guillaume Wisniewski et François Yvon

Résumé : Structured learning techniques, aimed at modeling structured objects such as labeled trees or strings, are computationally expensive. Many attempts have been made to reduce their complexity, either to speed up learning et inference, or to take richer dependencies into account. These attempts typically rely on approximate inference techniques and usually provide very little
theoretical guarantee regarding the optimality of the solutions they find.

In this work we study a new formulation of structured learning where inference is primarily viewed as an incremental process along which a solution is progressively computed. This framework generalizes several structured learning approaches. Building on the connections between this framework and reinforcement learning, we propose a theoretically sound method to learn to perform approximate inference. Experiments on four sequence labeling tasks show that our
approach is very competitive when compared to several strong baselines. Structured learning techniques, aimed at modeling structured objects such as labeled trees or strings, are computationally expensive. Many attempts have been made to reduce their complexity, either to speed up learning and inference, or to take richer dependencies into account. These attempts typically rely on approximate inference techniques and usually provide very little theoretical guarantee regarding the optimality of the solutions they find.

Dans la même collection

Sur le même thème