Section: New Results

Multiword expressions and statistical parsing

Participants : Sarah Beniamine, Marie-Hélène Candito, Benoît Sagot, Djamé Seddah.

Multi-word expressions recognition (MWE recognition) and syntactic parsing are two tasks that have been extensively investigated. Yet, systems combining both tasks have been rather rare. In particulat, works on parsing have tended to use training and test data with gold MWEs (generally with each MWE) merged into one token. In 2013, Djamé Seddah led the organization of the first shared task on statistical parsing Morphologically Rich Languages (SPMRL) [127] , hosted by the fourth SPMRL workshop. The primary goal of this shared task was to bring forward work on parsing morphologically ambiguous input in both dependency and constituency parsing, and to show the state of the art for MRLs. The shared task proposed a data set for 9 languages. The French part of this data set is particular, in that it uses a representation combining MWEs and syntax, which allows to investigate techniques for performing parsing and MWE recognition. A first system was proposed for the dependency parsing track of the Shared Task, in collaboration with Matthieu Constant (LIGM, Université Marne-la-Vallée) [74] . This work investigates pipeline and joint architecture for both tasks. In 2014, Marie Candito and Matthieu Constant continued that line of work [2] , focusing on using an alternative representation of syntactically regular MWEs, which captures their syntactic internal structure. The objective of such representation was two fold. First, it is well-known that the MWE status is not clear-cut, and that MWE status can hold due to syntactic and/or semantic criteria. In particular, syntactically regular MWEs exhibit various degrees of semantic non-compositionality. For such MWEs, an atomic representation fails to capture internal partial semantic composition, and also fails to take advantage of the internal syntactic regularity. Indeed, one hypothesis of this work was that augmenting the regularity of the syntactic representations could help parsing. The results of this work is that while this hypothesis could not be verified, the resulting system has comparable performance to that of previous works on this dataset, but it has the advantage of predicting both syntactic dependencies and the internal structure of MWEs, a crucial feature to capture the various degrees of semantic compositionality of MWEs.

In the same time, Sarah Beniamine and Benoît Sagot also investigated the use of internal regular structures for MWEs, yet for syntagmatic syntactic parsing. The objective is to guide a parser with predicted MWEs, while keeping a regular syntactic representation.