Section: New Results

Playing with DyALog-based parsers

Participants : Éric Villemonte de La Clergerie, Nicholas Parslow.

Éric de la Clergerie has continued the development of two DyALog-based parsers, namely DyALog-SR , a transition-based dependency parser, and FRMG, a wide-coverage French TAG based on an underlying meta-grammar.

The coverage of FRMG has been extended to cover more (rare) syntactic phenomena. A new conversion scheme has been added for the French version of the Universal Dependency Scheme. Preliminary evaluation experiments have been conducted on the French UD corpus, with both FRMG and DyALog-SR . FRMG has also been evaluated on the French SPMRL corpus, alone and with coupling with DyALog -sr

A new notion of secondary edges has been investigated in FRMG metagrammar and parser to provide additional dependency edges, helpful for understanding parsing outputs. In particular, secondary edges are used to denote controls between a verb and its hidden subject.

FRMG's disambiguation tuning is learned from CONLL-like treebanks using supervised learning method. We have conducted preliminary experiments to use unsupervised learning methods with observed accuracy gains between 1 to 1.5 points w.r.t. the no tuning case. However, trying to mix supervised and unsupervised methods have shown no significant gain w.r.t. the supervised case.

The hybridation of FRMG and DyALog-SR have been tried on a larger spectrum of treebanks.

FRMG has also been exploited during the Master internship of Nicholas Parslow about the use of NLP tools to provide feedback information and correlations on essays written by non-native French learners. In particular, the correction mechanism of FRMG has been extended to cover more cases of frequent errors and provide more explicit messages.