Sujet de thèse sans
financement (seuls les étudiants avec leur propre
financement peuvent choisir ces sujets:
étudiants étrangers avec une bourse de leur pays, normaliens,..)
PhD proposal without grant (only students with their own grant can be appointed on such PhD proposals, e.g. with foreign grants)
http://www.lirmm.fr/~retore/thesisproposals.html
Prerequisite / prérequis: advanced notions in logic / de
solides connaissances en logique
____________
The semantic analysis of texts, viewed as the mapping of a
sentence to one or several logical formulae requires a rich
semantic lexicon that associates each word with several typed
lambda terms describing its argumental structure (indicating,
roughly speaking “who does what to whom”) [LCG]. When one wishes
to take lexical meaning into account, i.e. to determine the
sense of each word in a given context, the lambda terms in the
semantic lexicon are even more complex: the sort of the entities
must be specified (as types) as well as the coercions that may
change the meaning of a word given its context. All this
information is represented as typed lambda terms that ensure the
semantics is computable. Using this lexical information, we can
infer from a sentence such as “Stephen started a book” that one
of its possible meanings is paraphrased as “John started reading
a book” (in many contexts, this is even the preferred meaning,
unless Stephen is a novelist like Stephen King).[MGL]
Grammars can be acquired automatically from
annotated corpora, and we do actually have a large French
categorial grammar integrated with a wide-coverage parser
[Grail], which provides semantic representations as logical
formulae, though without incorporating lexical meaning. The objective of the
PhD is to construct a semantic lexicon, i.e. to acquire the
typed lambda terms, and to integrate these into the logical
formulae automatically obtained. Rather than inferring the
lambda terms from annotated corpora we suggest to convert
information from an existing lexical network i.e. from
JeuxDeMots [JdM] into typed lambda terms.
A second objective will be to refine the
lexicon for application to a specific domain with a well-defined
terminology, e.g. for one or several of the online written
debates of the platform Dialoguea by Jean Sallantin, and to
evaluate the resulting system by testing it on some types of
entailments from this domain. These include entailments of the
type sketched above (from “start a book” infer “start reading a
book”) but also domain specific entailments. For example, in the
context of online debates, entailment allows us to determine (at
least in some clear cases) whether one debater adequately
paraphrases another or whether what seems a paraphrase
crucially shifts the meaning of some terms. Given the subtlety
of many of these inferences and the undecidability of
first-order logic, part of the problem consists of carving out a
problem which is at the same time interesting yet decidable in
enough of the cases commonly encountered.
_____________
[Grail] Wide-Coverage
French Syntax and Semantics using Grail Moot, Richard
Proceedings of Traitement Automatique des Langues Naturelles
(TALN) 2010
[JdM ] Chamberlain, J., Fort, K.,. Kruschwitz, U., Lafourcade,
M. & Poesio, M. Using Games to Create Language Resources:
Successes and Limitations of the Approach. Theory and
Applications of Natural Language Processing. Gurevych, Iryna;
Kim, Jungi (Eds.), Springer, ISBN 978-3-642-35084-9, 2013.
[LCG] Richard Moot et Christian Retoré The logic of categorial
grammars : a deductive account of natural language syntax
and semantics, LNCS 6850 Springer 2012.
[MGL] Christian Retoré (2014) The Montagovian Generative
Lexicon /\Ty_n: a Type Theoretical Framework for Natural
Language Semantics in R Matthes & A Schubert TYPES 2013
Postproceedings LIPICS 2014
http://dx.doi.org/10.4230/LIPIcs.TYPES.2013.202