| Question answering (QA) allows users to ask questions in natural language form and returns the exact answer directly. Compared with the traditional search engines, the query way of QA systems is more humane, and better to meet the users’needing to obtain the required information quickly and accurately. QA systems represent the develepment direction of the next generation of intelligent search engines.Question classification provides semantic restrictions and constraints for subsequent answer extraction and selection through determining the target answer type of the question. In detail, question classification can narrow the search range of candidate answers, and different answer selection strategies are chosen according to different types of questions, so it’s very significant for question classification to enhance the overall performance of QA systems.Question classification is similar to but harder than text classification. The objects of question classification are short sentences which contain less feature information. For the current question classification based on supervised learning, feature extraction is the key task owing to the fact that the more features are extracted, the higher classification accuracy obtained. However, there are two shortcomings to the existing feature extraction approaches:(1) Feature extraction depends on the Natural language processing(NLP) technology which is used to obtain useful features. This method not only exists the constraint of the NLP technology, but also can’t meet the real-time requirement of QA systems.(2) Feature extraction is emphasized but feature combination optimization is ignored. In existing work, features are combined only based on their importance to question classification. However, this method has some subjectivity, and affects the actual result of question classification.For Chinese question classification, above two shortcomings are more obvious. It’s more difficult to extract features from questions due to the fact that the Chinese NLP technology is not mature enough. Meanwhile, because of the variety and complexity of the Chinese language, the association between features of Chinese questions is much stronger, so it’s very important to optimize the combination of these features. This paper focuses on getting rich features and optimizing the feature combination, which can improve the performance of Chinese question classification. The detail work is as follows:(1) Aming at the difficulty of getting rich features for Chinese question classification, a feature model to integrate the basic features and bag_of_words(BOW) binding features is proposed. A new kind of question features-BOW binding features are generated automatically by binding part-of-speech, word meaning and other basic features with the BOW features, and then the basic features and BOW binding features are combined to obtain more efficient question feature set. The new feature model not only can be achieved easily and owns low processing cost, but also effectively make up the deficiency of the syntactic and semantic expression of basic features. Experimental results show that, when integrating BOW binding features to the basic features, the highest classification accuracy is comparable to the best known on the same kind of Chinese question set.(2) For the BOW binding operation that binding the part of speech, word meaning and other basic features with BOW separately mentioned in (1), the extension operation called multiple bag_of_words binding (MBWB) to bind those basic features with BOW together is proposed. Two types of potential question features-MBWB features and T_MBWB features are generated automatically by applying MBWB operation on BOW and trunk_BOW(T_BOW). The potential question features contained in basic features can be fully mined by MBWB operation, which also alleviates the situation that the existing feature extraction methods face the problem of language technology constraint to some extent. Experimental results show that question classification accuracy is improved obviously after adding MBWB features and T MBWB features to basic features.(3) Focus on the subjectivity of the method based on the importance analysis (IA), a new feature combination method is proposed based on the importance and inhibition analysis (IIA) which considering both the importance of a single feature itself and the inhibition between features. Experimental results show that the IIA method is more efficient for feature combination than the IA method on the whole.(4) Considering the efficiency of the IIA method for feature combination is low when a large number of features used, a new method i.e. difference and importance based feature combination (DIFC) is also proposed. The complementary of classifier is used for reference to measure the divisity between the features. The divisity degree is defined by considering both the correst and error classification result to the sample set. In addition, the complementary degree is defined by considering the divisity between the candidate features and the ones being combined, and the importance of the candidate features themselves. Experimental results show that the DIFC method is more flexible and efficient and gets higher accuracy compared with IIA and other feature combination method. |
