Section: New Results

Linear-time discriminant syntactico-semantic parsing

Participants : Benoit Crabbé, Maximin Coavoux, Rachel Bawden.

In this module we study efficient and accurate models of statistical phrase structure parsing. We focus on linear time lexicalized parsing algorithms (shift reduce) with approximations entailing linear time processing. The existing prototype involves a global discriminant parsing model of the large margin family (Perceptron,Mira, SVM) able to parse user defined structured input tokens [62] . Thus the model can take into account various sources of information for taking decisions such as word form, part of speech, morphology or semantic classes inter alia.

Our model has been generalized in a multilingual setting where we are among the state of the art systems and state of the art on some languages [23] . To our knowledge the parser is one of the fastest existing multilingual phrase structure parser. In order to ease model design for multilingual settings, we currently study efficient feature selection procedures for automating model adaptation to new languages.

We have also extended our model to continuous representations by means of deep learning methods. We currently have a neural network based decision procedure for parsing [22] . It involves both greedy search and beam based search techniques. Current work focuses on the design of dynamic oracles for improving greedy search procedures. This framework is currently tested in the multilingual setting too.

Further work involves to tackle the knowledge acquisition bottleneck problem by integrating either symbolic knowledge such as dictionaries or semi-supervised procedures for improving the formal representation of lexical dependencies in order to leverage data sparsity and estimation issues recurrent in lexicalized parsing.