| 摘要 |
PROBLEM TO BE SOLVED: To make it unnecessary to manually word-divide a large amount of texts for preparing the data for learning of a statistical language model, and to automatically learn the parameter of a language model from a learning text and a word dictionary. SOLUTION: A morpheme analysis mode or a parameter learning mode is judged (S1), and in the morpheme analysis mode (S1: NO), the morpheme analysis of an input learning text is operated (S2). In the parameter learning mode (S1: YES), the initial value of the parameter of a language model is estimated from the learning text and the word dictionary (S3), the morpheme analysis of a learning text is operated based on the appearance probability of a word applied by the language model (S4), and word division is operated, and the value of the parameter of the language model is re-estimated from the word-divided learning text (S5), this time parameter value is compared with the previous parameter value, and this processing is repeated until it is judged that the parameter is converged (S6: YES). |
