7th International Conference on Spoken Language Processing
September 16-20, 2002
The inevitable speech recognition errors make spoken dialogue systems not prevail in our daily lives. To reduce the impact of speech recognition errors, an error-tolerant language understanding model with confidence measuring is proposed in this paper. In this model, the exemplary concept sequences are used to provide the clues for detecting and recovering the errors. First, the hypothetical sentences are parsed to the hypothetical concept sequences. Then, the most matched pair of hypothetical and exemplary concept sequences is selected according to different kinds of scores, including the score of the edit operations (i.e., deleting, inserting and substituting concepts). The edit operation score is assessed according to the con- fidence level of the hypothetical concepts. Tested on cellular phone calls, the proposed model improves the precision rate of concepts from 65.09% to 76.32% and the recall rate from 64.21% to 69.13%, in comparison with the concept bigram model.
Bibliographic reference. Wang, Huei-Ming / Lin, Yi-Chung (2002): "Error-tolerant spoken language understanding with confidence measuring", In ICSLP-2002, 1625-1628.