Data Augmentation for Question Answering Using Transformer-based VAE with Negative Sampling

Abstract

In this paper, we propose a method to improve the accuracy of extracting appropriate question-answer pairs using generated questions with negative sampling. The base question-answering system that extracts similar questions for input queries is constructed on a Sentence-BERT model to carry out pairwised-ranking between questions of question-answer data and the input queries. The key issue of improving the question answering system is how we can prepare the enough size and variety of training examples. The Sentence-BERT model is trained on positive and negative pairs of extended questions generated by a Transformer-based Variational Autoencoder as well as human. Experimental results show that performance of retrieving appropriate questions for input queries is improved when the Sentence-BERT model is trained with the negative samples that are most similar to the positive examples.