Tree similarity measurement for classifying questions by syntactic structures

Abstract

Question classification plays a key role in question answering systems as the classification result will be useful for effectively locating correct answers. This paper addresses the problem of question classification by syntactic structure. To this end, questions are converted into parsed trees and each corresponding parsed tree is represented as a multi-dimensional sequence (MDS). Under this transformation from questions to MDSs, a new similarity measurement for comparing questions with MDS representations is presented. The new measurement, based on the all common subsequences, is proved to be a kernel, and can be computed in quadratic time. Experiments with kNN and SVM classifiers show that the proposed method is competitive in terms of classification accuracy and efficiency.