Syntactic Analysis 

Leader : Jacques Chauché


The objective is to use syntactical analysis for the study of semantic models of French mainly, with some incursions into other langages. The main two models with a semantic vector respresentation are tree transducers using Markov rewriting algorithms and Lambek pregroup grammars.

A majour tool came out of this research: a French text analyser called SYGMART. 
Syntactic analysis is not an end in itself. It's major contributions so far concern applications in data mining (See topic on text mining).

SYGFRAN has played an important role in the "EVALDA/EASY I et II " and "PASSAGE" evaluation campaigns.

Taking part in this area of research: Jacques Chauché, Anne Preller, Violaine Prince, Richard Terrat.



Lexical Semantics



Leader : Mathieu Lafourcade

A significant aspect of automatic language processing as a reseach field, is the increase of online dictionary as a resource. These resources, were very rare 15 years ago. They are now constantly enriched with lexical knowled base like WordNet, EuroWordNet, domain specific knownledge base, bilingual dictionaries between languages that had previously little relations between them (for example : French-Malay, Japanese-Arabic, etc ...).

The TAL team of LIRMM, working with GETALP of Grenoble, heavily contribute to theses resources, in several ways. One of them is the implementation of a serious game called JEUX DE MOTS which produces a lexical network enriched by anonymous contributors.

The team has also greatly contributed to representations of word meanings through vectors. Unlike traditionally used vectors, derived from the corpus, such as Salton vectors,  LSA or SVM vectors, the team chooses to develop an approach based on the very principles of the language.

- Semantic vector models or vector representations of semantic indexing in the manner of Roget: these vectors have the capacity to carry polysemy without disambiguation. They are of fixed size (the base never changes). Article explaining the basics of semantic vectors.

 - Conceptuel vectors: using semantic vectors as a basis, these vectors are enhanced with meanings coming from dictionaries (achieved by automatic processing of definitions) and usage in corpus. They are much more numerous and more discriminating: they do not convey multiple meanings. article on conceptual vectors


Text mining