An Effective Feature-Weighting Model for Question Classification

Abstract

Question classification is one of the most important sub- tasks in Question Answering systems. Now question tax- onomy is getting larger and more fine-grained for better answer generation. Many approaches to question classifi- cation have been proposed and achieve reasonable results. However, all previous approaches use certain learning al- gorithm to learn a classifier from binary feature vectors, extracted from small size of labeled examples. In this pa- per we propose a feature-weighting model which assigns different weights to features instead of simple binary val- ues. The main characteristic of this model is assigning more reasonable weight to features: these weights can be used to differentiate features each other according to their contri- bution to question classification. Furthermore, features are weighted depending on not only small labeled question col- lection but also large unlabeled question collection. Exper- imental results show that with this new feature-weighting model the SVM-based classifier outperforms the one with- out it to some extent.