Learning Tree Patterns for Syntactic Parsing

This paper presents a method for parsing Hungarian texts using a machine learning approach. The method collects the initial grammar for a learner from an annotated corpus with the help of tree shapes. The PGS algorithm, an improved version of the RGLearn algorithm, was developed and applied to learning tree patterns with various phrase types described by regular expressions. The method also calculates the probability values of the learned tree patterns. The syntactic parser of learned grammar using the Viterbi algorithm performs a quick search for finding the most probable derivation of a sentence. The results were built into an information extraction pipeline.