In this paper, a statistical framework for semantic parsing is described. The statistical model uses two information sources to disambiguate between rules: rule weights that capture vertical relationships in the parse tree, and concept n-grams that capture horizontal relationships. Rule design consists of simple local mapping rules that non-experts can write, and the rules are implemented as weighted finite state transducers. A general parser for context free grammars is implemented using a finite state machine library. Semantic decoding is implemented by recursively composing the rule transducer with the word-graph automaton produced from the speech recognizer. Detailed metrics for evaluating semantic parse accuracy are proposed. The parser is evaluated on the ATIS travel task with resulting precision and recall rates of over 95%. The proposed finite state transducer formulation allows the incorporation of rules and probabilities in a unified framework and the straightforward combination of acoustic, language, and understanding models.
Cite as: Potamianos, A., Kuo, H.-K.J. (2000) Statistical recursive finite state machine parsing for speech understanding. Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000), vol. 3, 510-513, doi: 10.21437/ICSLP.2000-584
@inproceedings{potamianos00b_icslp,
author={Alexandros Potamianos and Hong-Kwang J. Kuo},
title={{Statistical recursive finite state machine parsing for speech understanding}},
year=2000,
booktitle={Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000)},
pages={vol. 3, 510-513},
doi={10.21437/ICSLP.2000-584}
}