Sujet de these: PhD thesis offer in France/Learning from Post-Edition in Machine Translation / LIFL (Lille) and LIG (Grenoble)

[Thierry Hamon]

Date: Wed, 23 Jul 2014 23:26:34 +0200
From: Laurent Besacier <laurent.besacier at imag.fr>
Message-Id: <75FD7EE1-DFCB-4AC7-A8D4-4B3AA30BCA57 at imag.fr>


PhD thesis offer in France/ Learning from Post-Edition in Machine
Translation / LIFL (Lille) and LIG (Grenoble)


Contacts :   Olivier Pietquin : olivier.pietquin at univ-lille1.fr Laurent 
Besacier : laurent.besacier at imag.fr


Problem

Statistical Machine Translation (SMT) is the process by which texts are
automatically translated from a source language to a target language by
a machine that has been trained on corpora in both languages. Thanks to
progress in the training of SMT engines, machine translation has become
good enough so that it has become advantageous for translators to
post-edit machine outputs rather than translate from scratch. However,
current enhancement of SMT systems from human post-edition (PE) are
rather basic: the post-edited output is added to the training corpus and
the translation model and language model are re-trained, with no clear
view of how much has been improved and how much is left to be
improved. Moreover, the final PE result is the only feedback used:
available technologies do not take advantage of logged sequences of
post-edition actions, which inform on the cognitive processes of the
post-editor.

The proposed thesis aims at using the post-edition process as a
demonstration of how an expert translator modifies the SMT result to
produce a perfect translation. Learning from demonstration is an
emerging field in machine learning, mostly applied to robotics [1] that
will thus be explored further in the particular framework of SMT.

Topic of research

A novel approach to SMT training will be adopted in this thesis, i.e.
considering the post-edition process as a sequential decision making
process performed by human experts who should be imitated. This thesis’
first fundamental contribution to SMT will be to reformulate the problem
of post-edition in SMT as a sequential decision making problem
[4]. Indeed, the hypothesis selection and ranking process occurring in
an SMT system can be seen as an action selection strategy, choosing
after each post-edition step amongst a large number of actions (all
possible hypotheses and rankings). This strategy has to be modified
according to post-edition results arising sequentially and being
influenced by previous actions (hypothesis selection) of the system.

From this, SMT will be casted into an imitation learning problem, that
is learning from demonstrations made by an expert: post-edition results
can be seen as examples of what the system should do, again in a
sequential decision making process and not in a static one such as
supervised learning. Indeed, SMT decoding, whether it is based on
phrases or chunks, can be seen as a sequential decision making
process. The sequences of decisions taken by an expert during the
post-edition process can be seen as a target for the system, which will
try to imitate them in similar situations. To do so, we will extend the
work described in [2], that modelled semantic parsing as an Inverse
Reinforcement Learning (IRL) [3].

In addition, the question of automatically selecting the sentences that
should be used for post-edition and further learning will be addressed.
Especially, this will be studied under the active learning
paradigm. Large and diversified amounts of post-edited data, collected
in an industrial setting, will be made available for the research
project.


Profile

The applicants must hold an Engineering or a Master degree in
Computational Linguistics or computer science, preferably with
experience in the fields of statistical machine learning and/or natural
language processing. Good background in programming will also be
required. He/she will also be involved in a research project, funded by
the French National Agency for Research, involving 2 research labs (LIFL
in Lille and LIG in Grenoble) and a company (Lingua & Machina). For this
reason good English level is required (good command of French being a
plus). Finally effective communication skills in English, both written
and verbal are mandatory.

Context

The candidate will be hired by University Lille 1 in the framework of a
national research project. S/he will mainly be hosted in the SequeL (
Sequential Learning) team of the Laboratoire d’Informatique Fondamentale
de Lille (LIFL). SequeL is also a common team-project with INRIA
(national institute for research in computer science and mathematics)
and espe- cially the INRIA Lille - Nord Europe Center. The group
involves around 25 researchers working on sequential learning and is
internationally recognized. Lille is the largest city of the north of
France, a metropolis with 1 million inhabitants, with excellent train
connections to Brussels (30 min), Paris (1h) and London (1h30).

This thesis will be supervised in strong collaboration with the GETALP
team of Laboratoire d’Informatique de Grenoble (LIG), widely renowned
for its research on natural language and speech processing. Grenoble is
a high-tech city with 4 universities. It is located at the heart of the
Alps, in outstanding scientific and natural surroundings. It is 3h by
train from Paris ; 2h from Geneva ; 1h from Lyon ; 2h from Torino and is
less than 1h from Lyon international airport.

The PhD thesis will be co-supervised by Olivier Pietquin in Lille and
Laurent Besacier in Grenoble.

Contacts

Interviews will be held in Sept 2014. Meetings during Interspeech 2014
in Singapore can be also organized. For further info, please contact:

Olivier Pietquin : olivier.pietquin at univ-lille1.fr 
Laurent Besacier : laurent.besacier at imag.fr

References

 [1] Brenna D. Argall, Sonia Chernova, Manuela Veloso, and Brett
 Browning. A survey of robot learning from demonstration. Robotics and
 Autonomous Systems, 57(5):469–483, May 2009.

 [2] Gergely Neu and Csaba SzepesvÃąri. Training parsers by inverse
 reinforcement learning. Machine Learning, 77(2-3):303–337, 2009.

 [3] Andrew Y. Ng and Stuart J. Russell. Algorithms for inverse
 reinforcement learning. In Proceedings of the Seventeenth International
 Conference on Machine Learning, ICML ’00, pages 663–670, San Francisco,
 CA, USA, 2000. Morgan Kaufmann Publishers Inc.

 [4] Richard S. Sutton and Andrew G. Barto. Reinforcement Learning: An
 Introduction. The MIT Press, 3rd edition, March 1998.

-------------------------------------------------------------------------
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
-------------------------------------------------------------------------