Training Stochastic Grammars From Unlabelled Text Corpora

The paper describes various aspects and prac- ticalities of applying the "Hidden Markov" ap- proach to train parameters of regular and context- free stochastic grammars. The approach enables grammars to be trained from unlabelled text cor- pora, providing flexibility in the choice of syntac- tic categories and text domain. Part-of-speech tagging and parsing are discussed as applica- tions. Linguistic considerations can be used to de- velop constrained grammars, providing appropri- ate higher-order context for disamhiguation. Un- constrained grammars provide the opportunity to capture patterns that are not covered by a specific grammar. Experimental results are discussed for these alternatives.