Minimally supervised question classification on fine-grained taxonomies

This article presents a minimally supervised approach to question classification on fine-grained taxonomies. We have defined an algorithm that automatically obtains lists of weighted terms for each class in the taxonomy, thus identifying which terms are highly related to the classes and are highly discriminative between them. These lists have then been applied to the task of question classification. Our approach is based on the divergence of probability distributions of terms in plain text retrieved from the Web. A corpus of questions with which to train the classifier is not therefore necessary. As the system is based purely on statistical information, it does not require additional linguistic resources or tools. The experiments were performed on English questions and their Spanish translations. The results reveal that our system surpasses current supervised approaches in this task, obtaining a significant improvement in the experiments carried out.