Habilitation: Alexandre Allauzen, Modeles statistiques pour la traduction automatique

[Thierry Hamon]

Date: Thu, 16 Jan 2014 13:21:33 +0100
From: Alexandre Allauzen <allauzen at limsi.fr>
Message-ID: <52D7CECD.4040602 at limsi.fr>
X-url: http://www.limsi.fr/Pratique/acces/


J'ai le plaisir de vous inviter à ma soutenance d'Habilitation à Diriger
des Recherches intitulée "Modèles statistiques pour la traduction
automatique", ainsi qu'au pot qui suivra. La soutenance aura lieu le
jeudi 30 janvier 2014 à 14h30 dans la salle de conférence du LIMSI-CNRS.

Le jury est composé de :

Laurent Besacier, Professeur des Universités, Université J. Fourier,
  Grenoble 1
Alain Denise, Professeur des Universités, Université Paris-Sud
George Foster, Senior Research Officer, Conseil national de recherches 
  Canada
Hermann Ney, Professeur, Université d’Aix-la-Chapelle
Isabelle Tellier, Professeur des Universités, Université Paris 3- 
  Sorbonne Nouvelle
François Yvon, Professeur des Universités, Université Paris-Sud

Accès: pour venir au LIMSI, vous pouvez consulter la page web
http://www.limsi.fr/Pratique/acces/

--
Résumé:

The advent of Internet has deeply modified our communication habits and
especially our relationship with multilingualism: everyone needs to
access and broadcast information that are not in its native language,
and to electronically communicate with other Internet users who speak or
write in different languages. These different factors create a new
demand for translation services, that can only be realistically
fulfilled by machine translation tools.

With the exponential increase of electronic corpora along with the
development of statistical methods, tremendous progress has been made
over the last two decades in automatic translation of natural languages
(i.e Machine Translation). Nowadays, statistical machine translation
(SMT) can be considered as the most promising approach. However, given
the complexity of the task and its peculiarities, modelling and
estimation issues remain to be addressed and motivate a large part of my
work. My habilitation report is focused on 4 lines of my research
activities, each described in a chapter :

- The phrase-based SMT approach and more precisely its n-gram variant
  are described, along with an overview of my recurrent work on the
  development of state of the art SMT systems for international
  evaluation campaigns like WMT and IWSLT.

- The associations between a source and a target phrases and their
  statistics, which are at the heart of phrase-based SMT systems, rely
  on noisy information provided by word alignment models. This chapter
  investigates a model-based replacement of this important preliminary
  step, in which the posterior probabilities of the word alignment links
  are introduced in the construction of the translation models. The
  posterior distributions are estimated using a maximum-entropy
  classifier.

- Given the inventory of translation units, most SMT architectures
  follow a two-steps procedure in which several probabilistic models are
  first estimated independently, and then combined in a global linear
  model, the weights of which are chosen so as to maximize an overall
  translation quality criterion. In this line of research, an end-to-end
  discriminative model is introduced, in which all the parameters are
  learned discriminatively in a unified manner. The derivations of the
  translation process that defines reordering and segmentation
  operations are introduced as latent variables in an exponential model.

- Most statistical models for SMT consider words and phrases as events
  of discrete random variables. The resulting representations are very
  sparse and ignore relationships that may exist among these units. This
  lack of structure hinders the generalization and adaptation power of
  the resulting models. To remedy these issues, a continuous space
  translation model is introduced, in which words or phrases are
  projected in a continuous space, and represented by dense vectors of
  small dimensions.  The translation probabilities are then expressed as
  a smooth function of these representations. The projection and the
  probability estimation are jointly learned using a multi layer neural
  network.

The conclusion of the report summarizes my contributions and proposes
directions for future work. These directions are focused on word
reordering in SMT, continuous vector space models and weakly supervised
learning.

      Alexandre Allauzen
  Univ. Paris-Sud/LIMSI-CNRS
Tel : 01.69.85.80.41 (80.88)
   Bur : 29    LIMSI Bat. S
      allauzen at limsi.fr

-------------------------------------------------------------------------
Message diffuse par la liste Langage Naturel <LN at cines.fr>
Informations, abonnement : http://www.atala.org/article.php3?id_article=48
English version       : 
Archives                 : http://listserv.linguistlist.org/archives/ln.html
                                http://liste.cines.fr/info/ln

La liste LN est parrainee par l'ATALA (Association pour le Traitement
Automatique des Langues)
Information et adhesion  : http://www.atala.org/

ATALA décline toute responsabilité concernant le contenu des
messages diffusés sur la liste LN
-------------------------------------------------------------------------