Section: New Results

Playing with DyALog-based parsers

Participant : Éric Villemonte de La Clergerie.

Éric de la Clergerie has continued the development of DyALog-SR , a transition-based dependency parser running on top of DyALog and initiated in 2013 to participate to SPMRL'2013. Thanks to DyALog 's tabulation functionalities, this parser implements a dynamic programming algorithm to explore larger search space through the use of beams.

In order to participate to SemEval'14 Task on "broad coverage semantic dependency parsing", DyALog-SR was extended to handle non-connected dependency graphs rather than standard dependency trees. This was achieved by considering a richer set of transitions, besides the usual Shift and Reduce transitions. However, while working, this extended set of transitions was not ensuring the expected gains when using beams. The issue was finally solved after long investigations, with the identification of multiple causes. One of them was related to the fact that transition paths of various lengths may lead to a final state. In consequence, a noop transition was added to compensate on shorter paths.

A second axe of work was a thorough use of DyALog-SR over the French TreeBank (FTB) to compare its performances to those published for other parsers. By enriching its set of features and improving the update strategy of the perceptron-based statistical model of DyALog-SR , we were able to reach state-of-the-art results.

However, the best results were obtained by coupling DyALog-SR with FRMG, our large-coverage French grammar (derived from a meta-grammar). The results from FRMG were used as features to guide the statistical DyALog-SR parser. This innovative step proved to provide us with the best results published so far for the FTB (over 90% of Labeled Attachement Score [LAS] over the test part of the FTB) [41] .

The improvements of FRMG was pursued in 2014, at the level of the underlying meta-grammar (to extend its coverage over 96% on the FTB) but also by adapting the statistical models developed for DyALog-SR (in replacement of older and slower SQLite-based models).