Model learning methods, information extraction method, model learning device, information extraction apparatus, model learning program, information extraction program, and recording medium having recorded them program

摘要

PPROBLEM TO BE SOLVED: To provide a model learning method and an information extracting method that reduce influence of a recognition error. PSOLUTION: In a discrimination model, a terminal device generates model information during model learning by using recognition word string learning data which are given extraction object information and include an error and reference word string learning data which are given the extraction object information and include no error. Then, a recognition conviction degree identity is binarized and the model information is generated by using both a recognition word string and a reference word string. When information is extracted, recognition conviction degree identities showing whether respective recognized words recognized from input data are correct are imparted to the recognition word string. When a recognition conviction degree identity is larger than a certain threshold, the word is recognized "correctly" and when not, the word is recognized "incorrectly". Then the generated model information is used to impart the extraction object information to the recognition word string and words are extracted based upon the imparted extraction object information. PCOPYRIGHT: (C)2008,JPO&INPIT