The SPMRL shared task

Since several years, Djamé Seddah, together with Marie-Hélène Candito and more generally the whole Alpage team, has played a major role in setting up and animating an international network of researchers focusing on parsing morphologically rich languages (MRLs).

DyALog -SR

The SPMRL 2013 shared task was the opportunity to develop and test, with promising results, a simple beam-based shift-reduce dependency parser on top of the tabular logic programming system DYALOG. We used (Huang and Sagae, 2010) as the starting point for this work, in particular using the same simple arc-standard strategy for building projective dependency trees. The parser was also extended to handle ambiguous word lattices, with almost no loss w.r.t. disambiguated input, thanks to specific training, use of oracle segmentation, and large beams. We believe that this result is an interesting new one for shift-reduce parsing.

The current implementation scales correctly w.r.t. sentence length and, to a lesser extent, beam size. Nevertheless, for efficiency reasons, we plan to implement a simple C module for beam management to avoid the manipulation in DyALog of sorted lists. Interestingly, such a module, plus the already implemented model manager, should also be usable to speed up the disambiguation process of DyALog -based TAG parser FRMG (de La Clergerie, 2005a). Actually, these components could be integrated in a slow but on-going effort to add first-class probabilities (or weights) in DyALog , following the ideas of (Eisner and Filardo, 2011) or (Sato, 2008).

The Alpage-LIGM French parser

The second Alpage system that participated to the SPMRL shared task, although on French language only, was developed in collaboration with Mathieu Constant (LIGM), based on the Bonsai architecture. This system is made of several single statistical dependency parsing systems whose outputs are combined into a reparser. We use two types of single parsing architecture: (a) pipeline systems; (b)“joint” systems.

The pipeline systems first perform multi-word expression (MWE) analysis before parsing. The MWE analyzer merges recognized MWEs into single tokens and the parser is then applied on the sentences with this new tokenization. The parsing model is learned on a gold training set where all marked MWEs have been merged into single tokens. For evaluation, the merged MWEs appearing in the resulting parses are expanded, so that the tokens are exactly the same in gold and predicted parses.

The “joint” systems directly output dependency trees whose structure comply with the French dataset annotation scheme. Such trees contain not only syntactic dependencies, but also the grouping of tokens into MWEs, since the first component of an MWE bears dependencies to the subsequent components of the MWE with a specific label. At that stage, the only missing information is the POS of the MWEs, which we predict by applying a MWE tagger in a post-processing step.

This parsing system obtains the best results for French, both for overall parsing and for MWE recognition, using a reparsing architecture that combines several parsers, with both pipeline architecture (MWE recognition followed by parsing), and joint architecture (MWE recognition performed by the parser).