Learning Based Approaches for Vietnamese Question Classification Using Keywords Extraction from the Web

International Joint Conference on Natural Language Processing, pages 740–746, Nagoya, Japan, 14-18 October 2013. Learning Based Approaches for Vietnamese Question Classification Using Keywords Extraction from the Web Dang Hai Tran1, Cuong Xuan Chu1, Son Bao Pham1, Minh Le Nguyen2 1University of Engineering and Technology, Vietnam National University 2Japan Advanced Institute of Science and Technology 1{dangth,cuongcx_54,sonpb}@vnu.edu.vn 2nguyenml@jaist.ac.jp Abstract This paper presents our research on automatic question classification for Vietnamese using machine learning approaches. We have experimented with several machine learning algorithms utilizing two kinds of feature groups: bag-of-words and keywords. Our research focuses on two most important tasks which are corpus building and features extraction by crawling data from the Web to build a keyword corpus. The performance of our approach is promising where our system’s precision outperforms the state-of-the-art Tree Kernel approach (Collins and Duffy, 2001) on a Vietnamese question corpus. Keywords keyword collection, machine learning, Vietnamese question classification corpus. 1 Introduction Question Classification (QC) is a task that, given a question, maps it to one of the predefined k classes, which indicates a semantic constraint on the sought-after answer (Li and Roth, 2006). In a question answering system, before finding an answer of a question, the system has to iden- tify which category it is asking about, and this is the obligation of question classification. Then, based on the identified category, we can narrow the space of answers that we have to find. Let us consider some examples: • Q: "Ai là người phụ nữ đầu tiên hy sinh trong chiến tranh Việt Nam?" ("Who was the first woman killed in the Vietnam War?"), we ex- pect to know that the target of this question is a person, thus reducing the number of possi- ble answers significantly. • Q: Tại sao nắng màu vàng? (Why is the sunshine yellow?) indicates that this question wants to get information about reason, thus in next steps our system just concerns about reason answers space rather than human or number categories. The problem is that if the number of categories is more and the categories are more specific, the question answering system will spend more time classifying questions but it’s performance will be better. Let consider another example, in the next two questions, they are both asked to get informa- tion about location: • Q: Thành phố nào ở Canada có nhiều dân nhất? (What Canadian city has the largest population?) • Q: Đất nước nào trao tặng Mỹ tượng nữ thần tự do? (Which country gave New York the Statue of Liberty?) More particularly, we can see that the target of the first question is a city and the other one is a country, both city and country are locations. In this case, location is considered a coarse-grained class and city and country are fine-grained classes. Nat- urally, the more fine-grained classes we have, the more difficult it is to tell them apart. For hierar- chical categories, we adopted a two-level learning approach: first we solve the problem of classifying questions in …

Content truncated to 500 words