A generalized LR parser for text-to-speech synthesis

The development of a parser for a Norwegian text-to-speech system is reported. The generalized left-right (GLR) algorithm is applied, which is a generalization of the well-known LR algorithm for parsing computer languages. This paper briefly describes the GLR algorithm, the integration of a probabilistic scoring model, our implementation of the parser in C++, the attribute structures, the lexical interface, and the application of the parser to part-of-speech (POS) tagging for Norwegian. Applied to a small test set of about 4,000 words, this method correctly tags 96% of the known words, which is close to the performance of other POS-taggers trained on large text databases. 85% of the unknown words are tagged correctly, and the probability of choosing the wrong pronunciation of a word from the lexicon is less than 0.1%.