Probabilistic grammars are less adaptive to the user than are bigrams and trigrams. User-adaptation of grammar-rules is difficult, but an adaptable probabilistic dependence can be built into the way sentences are constructed within a given framework of rules. An optimising criterion using mutual information serves to link pairs of symbols within a derivation tree. The approach is data-driven, based on joint distributions of pairs of symbols obtained from training data. A simulation shows that the approach can provide an effective approximation to the true sentence probability. With adequate training this should lead to improved recognition performance. The dependence model can be built into a probabilistic parser with little loss of efficiency.
Cite as: Wright, J.H. (1991) Adaptation of grammar-based language models for continuous speech recognition. Proc. 2nd European Conference on Speech Communication and Technology (Eurospeech 1991), 203-206, doi: 10.21437/Eurospeech.1991-46
@inproceedings{wright91_eurospeech,
author={J. H. Wright},
title={{Adaptation of grammar-based language models for continuous speech recognition}},
year=1991,
booktitle={Proc. 2nd European Conference on Speech Communication and Technology (Eurospeech 1991)},
pages={203--206},
doi={10.21437/Eurospeech.1991-46}
}