Section: New Results

Task-based evaluation of syntactic lexica: coupling FRMG with various resources

Participants : Éric Villemonte de La Clergerie, Benoît Sagot.

The FRMG symbolic parser was used for comparing the performances of various syntactic lexicons as sources of information for parsing. The idea is to convert syntactic lexica other than the Lefff into the Lefff's format, i.e., turn them into Alexina lexicons, and then use the resulting lexica together with the FRMG grammar for producing several parsers. These parsers only differ by the lexical information they rely on. Preliminary results had already been obtained in 2009 [119] , but were restricted to one external lexicon, namely Lexicon-Grammar tables, and only to verbal entries (other entries were gathered from the Lefff when using Lexicon-Grammar-based verbal entries). However, conversion tools for other resources, such as Dicovalence [136] , had already been developed, in the context of the development and improvement of the Lefff. Moreover, the development of a new version of the Lefff verbal entries

Task-based evaluation results have been obtained on parsing with FRMG, showing that the Lefff performs better than both Lexique-Grammaire and Dicovalence (after conversion to the Alexina formalism) [48] , [49] . The new version of the Lefff, mentioned above, leads for now to lower results than the current version, but its results are better than with Lexique-Grammaire or Dicovalence data, despite a significant increase of the average amount of entries per lemma. These results are satisfying both because they show that the Lefff is a useful resource for symbolic parsing, but also because they illustrate the relevance of converting other resources into the Alexina formalism, in order to merge the valuable linguistic information they contain — as done in the last years for improving the Lefff [85] , [84] , [86] , [111] , [87] , [112] .