Intelligent Question Asking System Based On Maximum Entropy And Application In The Audit

With the development of information technology, the amount of information in auditfield increase drastically. The traditional software of computer-aided auditing graduallyexhibited the limitations of long audit cycle, low degree of intelligence when they wereprocessing these information. To overcome these limitations, we present a system ofIntelligent Question Asking in this thesis. It is different from the traditional audit techniqueswhich mainly check the report forms. Our system possesses high flexibility and it canascertain the doubtful points of audit.Our System consists of two main modules: one is automatic question generation (QG);the other is questionnaire constructing. QG takes texts in the field of social insurance as input.And we extract facts and rules from these texts to generate questions. One most importantmethod of QG is Maximum Entropy. First, we analysis the frequently used types and its focustypes of Chinese question. After artificially annotating representative policy texts in the fieldof social insurance, we employ these corpuses to train the Maximum Entropy Model. Whenthe parameters of the model are determined, it is used to annotate the Chinese sentenceswhich are going to generate questions. At last, we use QG algorithm to produce questions.These questions are engaged to construct a questionnaire referencing techniques of criminalpsychological test. Finally we provide principles and methods in questionnaire constructing.The system we present in the thesis is able to assist audit. Its features are as follows. Itcan generate questions according to the inputs and can construct questionnaires from thesequestions referencing the constructing techniques of criminal psychological testing. By askingquestions, we can judge the staff whether or not breaking the law. So we can discover thedoubtful points of audit. At last, the purpose of audit is to be achieved. It offers an efficient,intelligent means of audit. The experiments show that using Maximum Entropy methodgenerating questions and applying them to assist audit is feasible.