Exploitation of high level syntactic models for Natural Language Processing (ALPAGE)

Type de poste : Post-doctorant
Lieu de travail : Rocquencourt
Thème de recherche : Systèmes symboliques
Projet : ALPAGE
 
 
Environnement

INRIA project-team ALPAGE works in the field of Natural Language
Processing [NLP]. Guided by theoretical results in linguistics and
formal language theory, we develop various tools (such as parsers) and
French linguistic resources (grammars, lexicon) and apply them to
process corpora.
 
Missions

One of the tools developed by ALPAGE is the DyALog system, which may
be used to build parsers for various grammatical formalisms, including
Tree Adjoining Grammars (TAGs). DyALog implements theoretical results
issued from our works on several extensions of Push-Down Automata
(PDA).

The first objective of this proposal is to complete the DyALOg
compilation phase to integrate recent results on Thread
Automata. These automata generalize all the kinds of automata
currently covered by DyALog, hence suggesting a more uniform
processing of all the grammatical formalisms known to DyALog. These
automata also open the way for a complete handling of Mildly Context
Sensitive [MCS] formalisms, a large and linguistically motivated class
of formalisms that, in particular, includes local Multi-Component TAGs
(MC-TAGs) . The basic idea of MCTAGs is to replace grammatical
productions given as elementary parse trees by productions given as
sets of coordinated elementary parse trees, providing more flexibility
to handle deep extraction phenomena and scoping phenomena.

Once done the integration of Thread Automata within DyALog, the second
objective is the exploration of MC-TAGs, in relation with the notion
of meta-grammars. Meta-grammars provide an abstract level of syntactic
description that is followed by a phase of expansion towards a target
formalism like TAGs, and, in the future, MC-TAGs. The use of
meta-grammars aims to facilitate the development of grammars for new
formalisms. We already have a French meta-grammar FRMG whose TAG
expansion has been tested at large scale and that ensures a good
linguistic coverage. However, the evolution of FRMG is currently
hindered by the limitations of TAGs, motivating the switch to MC-TAGs
through the meta-grammar level.

The proposal includes the following tasks:
- Implementing MCTAGs into DyALog, through the notion of Thread
  Automata.
- Extending MGCOMP in order to produce MCTAGs rather than simply
  TAGs. The changes are minimal.
- Transforming and extend the current French MG FRMG to handle some
  phenomena through MCTAG.
- Running some experiments on corpus with the resulting parser.

Links:

ALPAGE: http://alpage.inria.fr/
references: http://alpage.inria.fr/biblio
with keywords thread, DyALog, metagrammar demo FRMG and DyALog: 
http://alpage.inria.fr/parserdemo
 
Compétences et Profil

Competences in parsing for NLP (algorithmic, techniques, formalisms).
Good practice of a Logic Programming language, like Prolog.
Familiarity with Linux environments.
 
Informations complémentaires

http://alpage.inria.fr/stages.en.html

Contact :
Eric.De_La_Clergerie@inria.fr