Accessibility setting

Section: New Results

Deep syntactic parsing

Participants : Corentin Ribeyre, Marie-Hélène Candito.

Syntax plays an important role in the task of predicting the semantic structure of a sentence. But syntactic phenomena such as alternations, control and raising tend to obfuscate the relation between syntax and semantics. We have investigated how to predict the semantic structure of a sentence, encoded using the FrameNet model, taking advantage of deeper syntactic information than what is usually used. This deep syntactic representation abstracts away from purely syntactic phenomena and proposes a structural organization of the sentence that is closer to the semantic representation, by normalising the syntactic paths between a verb and its arguments. This reduces the variety of the syntactic realization of semantic roles, as shown by a decrease of the entropy of the syntactic paths of a given role.

Experiments conducted on a French corpus annotated with semantic frames showed that a FrameNet semantic parser reaches better performances with such a deep syntactic information [31]. For instance, switching from surface to deep syntactic information leads to a significant gain in FrameNet role identification, especially when this information is predicted (rather than reference information): $+5.1$ points ($56.7$ to $61.7$) on all triggers (In the sense of FrameNet, i.e. predicative lexial units, which should be assigned a frame.) and $+6.7$ points ($61.3$ to $68.0$) on verbal triggers only. These results clearly show the benefit of using deep syntactic features.