Extracting Compact Sets of Features for Question Classification in Cognitive Systems: A Comparative Study

Abstract

Question Classification is one of the key tasks of Cognitive Systems based on the Question Answering paradigm. It aims at identifying the type of the possible answer for a question expressed in natural language. Machine learning techniques are typically employed for this task, and exploit a high number of features extracted from labelled questions of benchmark training sets in order to achieve good classification results. However, the high dimensionality of the feature space often limits the possibility of applying more efficient classification approaches, due to high training costs. In this work, more compact sets of lexical and syntactic features are proposed to distinguish question classes. In particular, the widely used unigrams are substituted with a smaller number of features, extracted by modifying typical Natural Language Processing procedures for question analysis. The accuracy values gained on a benchmark dataset by using these different sets of features are compared among them and with the state-of-the-art, taking into account the required complexity at the same time. The new sets of extracted features show a good trade-off between accuracy and complexity.