Section: New Results

Graph-based approaches for deep-syntatic and semantic parsing

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

With most state-of-the-art statistical parsers routinely crossing a ninety percent performance plateau in capturing tree structures, the question of what next crucially arises. Most of the structures used to train current parsing models are degraded versions of a more informative data set: the Wall Street journal section of the Penn treebank ( [91] ) which is often stripped from its richer set of annotations (i.e. traces and functional labels are removed), while, for reasons of efficiency and availability, projective dependency trees are often given preference over richer graph structures [96] , [107] . This led to the emergence of surface syntax-based parsers [70] , [97] , [100] whose output cannot by itself be used to extract full-fledged predicate argument-structures. For example, control verb constructions, it-cleft structures, argument sharing in ellipsis coordination, etc. are among the phenomena requiring a graph to be properly accounted for. The dichotomy between what can usually be parsed with high accuracy and what lies in the deeper syntactic description has initiated a line of research devoted to closing the gap between surface syntax and richer structures.

At Alpage, we built our work on the widely known transition-based parsing approach [95] , which is state-of-the-art to parse surfacic syntatic trees [141] . Shift-reduce transition-based parsers essentially rely on configurations formed of a stack and a buffer, with stack transitions used to move from a configuration to the next one, until reaching a final configuration.