We can reduce the error rate of continuous speech recognition, if we know the intention of the next utterance. Many researchers have been building the dialogue models which can predict the intention of the next uttereance. But the modeling of natural dialogue is difficult, and the number of intentions predicted with a dialogue model which prevents us from increasing the recognition rate can be large.
This thesis proposes three intention analysis methods which can analyze the result of speech recognition to increase the recognition rate. The first method uses the relations between intentions and words. The second method uses MIG(Markov intention graph) which is based on the relations between intentions and word transitions. The third method uses IDT(intention decision table) which is based on the relations between intentions and the components of the sentences. All methods are based on sentence patterns and can be applied not only to text data but also to speech data. We can also reduce the error rate from the result of speech recognition using the intention dependent language model.
Among the poposed three methods, MIG and IDT give good results in intention analysis. We achieved 96% accuracy in the text data and 87% in speech data by using them. We can reduce the speech recognition error rate by 22% using the analyzed intentions. These methods contribute not only to increasing the speech recognition performance, but also to the design of dialogue models with automatic analysis of dialogue data.
