Comparison between the inside-outside algorithm and the viterbi algorithm for stochastic context-free grammars

Abstract

The most popular algorithms for the estimation of the probabilities of a context-free grammar are the Inside-Outside algorithm and the Viterbi algorithm, which are Maximum Likelihood approaches. The difference between the logarithm of the likelihood of a string and the logarithm of the likelihood of the most probable parse of a string is upper bounded linearly by the length of the string and the logarithm of the number of non-terminal symbols. However, this theoretical bound is too pessimistic. For this reason, an experimental work to show the behaviour of the two functions in practical cases is necessary.