Séminaire de Recherche en Linguistique

A real-life parsing system should comprise the recognition of
multi-word expressions (MWEs) for two main reasons.
First, downstream semantic-oriented applications need some marking in
order to distinguish between regular semantic composition and the
typical non-compositionality of MWEs. Second, MWE information is
supposed to help parsing.
In this talk, we will present various strategies to predict both
syntactic dependencies and MWEs, testing them on the dependency
version of the French Treebank (Abeillé and Barrier, 2004), as
instantiated in the SPMRL Shared Task dataset for French (Seddah et
al., 2013).
We will focus on (i) enhancing a pure joint system for both kinds of
prediction and (ii) the use of an alternative representation of
syntactically regular MWEs, which captures their internal syntactic
structure. We derive from it a system that predicts both syntactic
dependencies and the internal structure of MWEs, which can be useful
for capturing the various degrees of compositionality that MWEs
exhibit.