Imbalanced Question Classification Using Generative Prototypes

Abstract

A practical question-answering (QA) system typically categorizes a new question into the frequently asked questions (FAQs) and returns the corresponding answer. Having imbalanced FAQs data is actually prevalent in general. This paper proposes a meta-learning method for imbalanced question classification. The basic idea is to generate virtual training data for zero-shot questions and then construct question prototypes for training a question classifier, thereby relieving the problem of data imbalance and improving performance of question classifier. Experiments show that the proposed method improves the overall classification performance both for English and Chinese QA tasks. Especially, the classification performance of zero annotated questions increased significantly, and the generative prototypes has minute impact on the performance of large annotated question test set.