TRAINING QUESTION CLASSIFIERS USING SUPPORT VECTOR MACHINES

Abstract

Open-domain question answering and story comprehension has great importance in natural language processing. In order to respond correctly to a free form factual question, one needs to understand a question to a level that allows determining the constraints to be imposed on the possible answers. These constraints not only include a semantic classification of the sought after answer, but also suggest using different strategies when looking for and verifying candidate answers. In fact, all the open domain question answering systems include a question classifier module. By definition, Question classification is a process that maps a given question to one of the k classes, which provide semantic constraints on the sought-after answer. The accuracy of the question classifier is very important to the overall performance of the question answering system. Previous research work has shown interest in using machine learning approaches for question classification. In this dissertation we have trained_ question classifiers with reasonably high accuracies using Support Vector Machines (SVM). The question classifier has been trained using a combination of several feature sets like words, n-grams, part-of-speech tags and syntactic tree information. We have also shown an increase in question classifier accuracy with increase in training dataset size. The biggest advantage of our dissertation work on question classification is that it gives high accuracy even when it is trained by just using surface text features like words and n-grams. The system runs on a Pentium III, 1 GHz Linux machine with 128MB of RAM. The language used for code development is Perl. LIBSVM, a library for SVM, is used for classification.