CFG-based language models have become popular over the last few years, especially for commercial applications, and there is growing interest in creating complex CFG-based models for mixed initiative systems. On general grounds, it is attractive to attempt to compile these models from domain-independent descriptions written in high-level formalisms such as unification grammar. Experience to date however suggests that compilation from complex unification grammars to CFG has poor scalability properties. We argue that it is possible to attack this problem by first specialising the domain-independent grammar against a corpus using Explanation Based Learning. We describe experiments carried out on a medium vocabulary command and control task, which suggest that language models derived from specialised grammars have much better scalability properties, and also deliver significantly improved run-time performance.
Cite as: Rayner, M., Hockey, B.A., Dowding, J. (2002) Grammar specialisation meets language modelling. Proc. 7th International Conference on Spoken Language Processing (ICSLP 2002), 913-916, doi: 10.21437/ICSLP.2002-306
@inproceedings{rayner02_icslp,
author={Manny Rayner and Beth Ann Hockey and John Dowding},
title={{Grammar specialisation meets language modelling}},
year=2002,
booktitle={Proc. 7th International Conference on Spoken Language Processing (ICSLP 2002)},
pages={913--916},
doi={10.21437/ICSLP.2002-306}
}