Questions Feature Extraction and Semi-supervised Classification Based on Terms Relevance

Question classification, an important component of question answering systems, has a direct impact on the answer extraction accuracy. In this paper, a question classification method is proposed by combined the question feature extracting of term relevance with semi-supervised classification. In detail, the method extracts structure terms in interrogative sentences as the feature space through statistical means, and calculates the relevance among terms by literal similarity method, besides, feature vectors of question classification are obtained by using term similarity relationship to build the questions' feature value in feature space. And then, utilizing unlabeled samples classify questions with the help of Co-training style and semi-supervised learning algorithm. Experimented on 20,000 questions in Yunnan tourism domain, the results show that more remarkable effects have been achieved by adopting the method above. The classification accuracy rate reaches 82.34%, which is higher than the TFIDF feature extraction methods and supervised learning methods by 15.4 percentage points and 1.4 percentage points separately.