PURPOSE:To preferentially delete common starting information whose contribution degree is low and to execute efficient learning compared with a case when sifting is not executed by checking a condition in accordance with a rule which is previously provided for extracted common starting information so as to rank them and executing sifting by means of the contribution degree for improving a conversion ratio. CONSTITUTION:Modification expressions stored in a modification analysis result storage part 14 are sequentially read by a modification expression classification part 15, and they are outputted to a modifying destination fuzziness discrimination part 16 and a modifying destination adjacency discrimination part 17. Then, they are classified based on results from the two discrimination parts. The modifying destination fuzziness discrimination part 16 discriminates whether a modifying destination is uniquely decided or not, and the modifying destination adjacency discrimination part 17 discriminates whether words in a modification relation are adjacent or not. Information for the discriminated result is returned to the modification expression classification part 15. The classified modification expressions are read by a common starting learning storage control part 7 and they are added to a common starting information storage part 6 as a pair of a modifier-a modified word on the modification expression belonging to previously decided classification.
展开▼