Section: New Results

Advances in symbolic parsing with DyALog /FRMG

Participant : Éric Villemonte de La Clergerie.

Within the team is developed a wide-coverage French meta-grammar (FRMG) and a efficient hybrid TAG/TIG parser based on the DyALog logic programming environment [127] and on the Lefff morphological and syntactic lexicon [118] . It relies on the notion of factorized grammar, themselves generated from a representation that lies at a higher level of abstraction, named Meta-Grammars  [129] . At that level, linguistic generalizations can be expressed, which in turn makes it possible to transfer meta-grammars from one language to a closely related one. The hybrid TAG/TIG parser generator itself implements all kinds of parsing optimizations: lexicalization (in particular via hypertags), left-corner guiding, top/bottom feature analysis, TIG analysis (with multiple adjoining), and others.

Éric de La Clergerie has continued to improve the coverage, quality and efficiency of the French meta-grammar FRMG. On the EasyDev corpus (around 4000 sentences), parsing times have improved over 2011 from an average of 1.03s per sentence to 0.28s, coverage (in terms of sentences with full parses) has improved from 72.5% to 82.60%, and accuray (in terms of f-mesure over relations) from 64.54% to 68.28%.

A part of the accuracy gains comes from the addition of a new output format for FRMG, namely the CONLL format, allowing us to use the CONLL-based dependency version of the French Treebank (around 12K sentences) for training and evaluation. We also used new machine learning techniques to improve FRMG's disambiguation algorithm, allowing us to combine heuristic based disambiguation rules (with manually provided weights) with more standard parsing features associated with automatically learned weights. More precisely, the idea was to study the efficiency of the disambiguation rules over the French treebank and to favor (resp. penalize) well-working (resp. bad working) rules by adjusting their weight, taking into account additional (and more standard) features. Using these techniques, on ftb6_3 test part, FRMG improved from a base accuracy of 82.31% (in terms of CONLL Labeled Attachment Score) to 84.54%. These gains resulting from a training over the French TreeBank have also been observed (with however a lesser impact) on the EasyDev corpus (using a different format and using a different evaluation metric).