Antoine Rozenknop

Discriminative Models of SCFG and STSG

Standard stochastic grammars use generative probabilistic models, focusing on rewriting probabilities conditioned by the symbol to be rewritten. Among several other undesired behaviors, such grammars tend to give penalty to longer derivations of the same input, which is a drawback when they are used for analysis (rather than generation). In this contribution, we propose a novel non-generative probabilistic model for both Stochastic Context-Free Grammars (SCFGs) and Stochastic Tree-Substitution Grammars (STSGs), in which the probabilities are conditioned by the leaves (i.e. the input symbols) rather than by the root of the parse tree. Both theoretical and experimental improvements of these new models are presented.

Gibbsian Context-Free Grammar for Parsing

Probabilistic Context-Free Grammars can be used for speech recognition or syntactic analysis thanks to especially efficient algorithms. In this paper, we propose an instantiation of such a grammar, which mathematical properties are intuitively more suitable for those tasks than SCFGs (Stochastic CFG), without requiring specific analysis algorithms. Results on the Susanne text show that up to 33% of analysis errors made by a SCFG can be avoided with this model.