Section: New Results

French Deep Syntactic Dependency Parsing

Participants : Corentin Ribeyre, Djamé Seddah, Éric Villemonte de La Clergerie, Marie Candito.

At Alpage, we used two distinct but complementary approaches to parse and produce deep syntactic dependency graphs from the DeepSequoia and the DeepFTB (crossref here). The first one was developped by using OGRE [87] , [86] , a graph rewriting system (crossref here). We developed a set of rewriting rules to transform surfacic syntactic dependency trees into deep syntactic dependency graphs, then we applied this set of rules on previously parsed surfacic trees. Those trees were produced using up to three different surfacic syntactic parsers: FRMG [109] , DyALog-SR [109] and Mate [47] . The results were convincing and on par with what we got on English.

The second approach was based on the work made last year regarding the English broad-coverage semantic dependency parsing. We reused our two graph parsers (the first one is based on a previous work on DAG parsing [89] and the second one on the FRMG surfacic syntactic parser [109] ) to parse the same graphs. As we previously have shown on English, the use of a mix of syntactic features (tree fragments from a constituent syntactic parser [80] , dependencies from a syntactic parser [47] , elementary spinal trees using a spine grammar [102] , etc.) improve our results. Our intuition is that syntax and semantic are not independent of each other and using syntax could improve semantic parsing. Finally, we extended a dual-decomposition third-order graph parser [76] to incorporate our syntactic feature set and we were able to reach the best performances to this day on the task for both English [28] and French (Ribeyre et al, to appear).