Section: New Results

Linear-time discriminant syntactico-semantic parsing

Participants : Benoit Crabbé, Maximin Coavoux, Djamé Seddah.

In this module we study efficient and accurate models of statistical phrase structure parsing. We focus on linear time lexicalized parsing algorithms (shift reduce, left corner) with approximations entailing linear time processing. The existing prototype involves a global discriminant parsing model of the large margin family (Perceptron,Mira, SVM avatars) able to parse user defined structured input tokens [23] . 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 participation to the SPRML 2014 shared task on parsing morphologically rich languages has been a first step towards testing our model in a multilingual setting where we were among the state of the art systems and state of the art on some languages such as Polish. To our knowledge the parser is one of the fastest existing multilingual parser worldwide (4000   8000 tokens/sec.). In order to ease model design for multilingual settings, we currently study efficient feature selection procedures for automating model adaptation to new languages.

The ongoing investigation aims to integrate continuous semantic representations into the model such as word embeddings in order to leverage data sparsity and estimation issues recurrent in lexicalized parsing. To this end we study neural-network-based architectures for structured phrase structure parsing.