This paper describes an approach to the significant improvement of the sentence accuracy before the linguistic analysis of
spontaneous pronounced queries to the “natural speaking” dialogue information retrieval system by using language models of
several kinds. The results of an application of language models of three different types (n-gram, polygram and permugram ones) to the output of a standard word recognizer are presented after the description of the
theoretical background of all used language models.
The work presented in this paper was partly supported by the Copernicus Joint Research Project COP-94/1634 SQEL and by the
Czech Ministry of Education under contract number MSM 2352 00005. The authors would also like to sincerely thank the colleagues
from the University of Erlangen for valuable suggestions and discussions.