Generating Diverse and Consistent QA pairs from Contexts with Information-Maximizing Hierarchical Conditional VAEs

One of the most crucial challenges in question answering (QA) is the scarcity\nof labeled data, since it is costly to obtain question-answer (QA) pairs for a\ntarget text domain with human annotation. An alternative approach to tackle the\nproblem is to use automatically generated QA pairs from either the problem\ncontext or from large amount of unstructured texts (e.g. Wikipedia). In this\nwork, we propose a hierarchical conditional variational autoencoder (HCVAE) for\ngenerating QA pairs given unstructured texts as contexts, while maximizing the\nmutual information between generated QA pairs to ensure their consistency. We\nvalidate our Information Maximizing Hierarchical Conditional Variational\nAutoEncoder (Info-HCVAE) on several benchmark datasets by evaluating the\nperformance of the QA model (BERT-base) using only the generated QA pairs\n(QA-based evaluation) or by using both the generated and human-labeled pairs\n(semi-supervised learning) for training, against state-of-the-art baseline\nmodels. The results show that our model obtains impressive performance gains\nover all baselines on both tasks, using only a fraction of data for training.\n