| 摘要 |
<p>PURPOSE: To improve the efficiency of recognition at a character processor based on a neural network. CONSTITUTION: It is discriminated whether a current character from the neural network is the final character of a word or not (202). When it is not the final character, that current character is compared with a character at the correspondent position of a dictionary item, and a candidate word list is prepared (205). At such a time, a reliability level is stored for each character. When the final character of the word is identified, the neural network word of peak probability is compared with the candidate word list (209 and 210) and when there is a coincident word, that neural network word is selected as a text word (211). When there is no coincident word, the difference between the candidate word of peak probability in the candidate word list and the neural network word is calculated (213) and when this difference is less than a prescribed value, the neural network word is selected as the text word (215) but in the other case, the candidate word of peak probability is selected (216).</p> |
