Third International Conference on Spoken Language Processing (ICSLP 94)
This paper describes implementation issues of the HMM-LR parser and preliminary evaluation results for parsing speed on a continuous HMM recognition system. The HMM-LR parser is based on an LR parsing technique which can use context-free grammars, and thought as one of possible parsing techniques which would realise real time processing. However, the current problem is that the total recognition system still needs much memory space and much processing time. We have implemented the HMM-LR parser to expand our existing speech recognition system to a sentence understanding system. In this paper, we present software implementation results and propose a couple of new techniques for reducing parsing time and memory space. The new ideas proposed are based on the encoding scheme of the LR table entries, table compression methods, and a novel beam search control. We have found that fast parsing speed has been realised for a textual input consisting of 1000 sentences. However, preliminary evaluations on the total system suggested the necessity of more intensive improvement on the parsing algorithm and implementation.
Bibliographic reference. Wallerstein, Frank O. / Amano, Akio / Hataoka, Nobuo (1994): "Implementation issues and parsing speed evaluation of HMM-LR parser", In ICSLP-1994, 9-12.