Section: New Results

English Broad-coverage Semantic Dependency Parsing

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

We successfully tested our graph-based approach described in Section  6.7 on a shared task on broad-coverage semantic dependency parsing part of the International Workshop on Semantic Evaluation (SemEval 2014, [99] ). We were given three resources, which constitute parallel semantic annotations over the same common text (the Penn Treebank (PTB), [91] ). The first one is part of the tectogrammatical layer of the Prague Czech-English Dependency Treebank, the second one is the reduction of the Minimal Recursion Semantics, available through the HPSG annotation of the PTB, into bi-lexical dependencies [82] . Finally, the third one is the predicate-argument structures extracted from the Enju Parser [131] . The shared task consisted of two tracks: a closed one where we needed to use these three resources only and an open one, where we could use whatever we needed to produce the best semantic representations.

At Alpage, we developed two semantic parsers: The first one is based on a previous work on DAG parsing [107] and the second one on the FRMG surfacic syntactic parser [133] . We use two parsers to assess the validity of our approach. The top performing models we submitted used a mix of syntactic features (tree fragments from a constituent syntactic parser [100] , dependencies from a syntactic parser [58] , elementary spinal trees using a spine grammar [126] , etc.) to improve our results. Our intuition is that syntax and semantic are not independent of each other and using syntax could improve semantic parsing. Our systems performs well and were able to compete with the top performers. Those systems, as well as the software needed to parse these new data sets, are already available.