Reducing the computational complexity for inferring stochastic context-free grammar rules from example text

Abstract

Learning context-free grammar rules from example text is a difficult task. Existing algorithms exhibit a cubic relationship between the number of non-terminal symbols and the computational complexity which has so far prevented the use of such models for applications with more then, say, 30 non-terminal symbols. The paper introduces a mathematical technique which reduces the complexity to a quadratic one, allowing the training of grammars with far more non-terminal symbols than was previously possible. Different versions of the algorithm are experimentally compared and the difference in order is verified. >