Integrating Selectional Constraints Satisfaction and Subcategorization Frame Selection in a Dependency Parser

Abstract

Dynamic Programming is a natural framework for projective dependency parsing. But, for efficiency reasons, it makes very strict independence hypotheses that prevent to adequately model important linguistic phenomena among which Subcategorization Frame Selection or Selectional Constraints Satisfaction. Integer Linear Programming, on another hand, is a general optimization method that allows to select an optimal set of linguistically relevant patterns in a sentence, such as Subcategorization Frames or Selectional Constraints. But is not a very natural framework for parsing. We propose in this paper an original way to combine both processes in order to produce parses that take into account Subcategorization Frames and Selectional Constraints as well as general syntactic constraints. Experiments on a French corpus showed a decrease of labeled error rate of 15,5% with respect to a state of the art parser.