Section: New Results

Comparing lexical resources for parsing: Lexicon-Grammar vs. Lefff

Participants : Benoît Sagot, Éric Villemonte de La Clergerie.

Lexicon-Grammar tables are currently one of the major sources of syntactic lexical information for the French language. Moreover, several Lexicon-Grammar tables exist for other languages. However, current tables suffer from various types of inconsistency and incompleteness. In particular, defining features are not represented in the tables. To remedy this situation, tables of classes are being developed at IGM (Université Paris-Est) for each category, and notably for verbs, which associate the set of their defining features with each class. Preliminary results of this long-term effort allowed Benoît Sagot, in collaboration with Elsa Tolone from IGM, to convert verb tables into the Alexina format, the format of the Lefff , hence turning it into a lexicon usable by the FRMG parser, named LgLex [44] , [43] , [49] .

This allowed Éric de La Clergerie and Benoît Sagot to build a variant of FRMG that doesn't uses the Lefff , but that uses a lexicon derived from the Lefff in which lexical entries for standard verbs are replaced by those from LgLex. We evaluated both FRMG variants on the EASy corpus with the EASy metrics. For now, results are still slightly better with the original Lefff . However, in this process, we identified many possible improvements, in particular in the tables themselves and in the conversion process. Moreover, a very promising complementarity between both resources has been found (the set of parsable sentences strongly vary w.r.t. the lexicon used), which confirms Alpage's approach to lexical development that emphasizes, among other techniques, the conversion, comparison and merging of existing lexical resources.