Section: New Results

Linear time constituent parser

Participant : Benoît Crabbé.

We have designed an efficient and accurate lexicalized LR inspired discriminative parsing algorithm that recasts some current advances in dependency parsing to the constituency setting. We specifically designed and evaluated a Graph Structured Stack-based parser (Huang et al. 2010) using some additional specific approximate inference techniques such as the max violation update for the perceptron (Huang et al. 2012) . By contrast with dependency parsing however, lexicalized constituent parsing raises some additional correctness issues that motivate the explicit use of an LR automata instead of a simpler shift reduce framework.

The parsing model is linear in time and has been evaluated on French data, where it turns out to be state of the art on SPMRL 2013 datasets [29] both in time and in accuracy. The parsing framework has been designed to be further extended with compositional semantic representations and allows in principle an easy integration of ressources — such as those developped in the team — considered to be important for parsing morphologically rich languages.