Zero-shot Question Generation: Accelerate the Development of Domain-specific Dialogue System

Neural-based approaches to question generation from structured data in specific domain are data-hungry, which makes it difficult to apply these methods in the real world. In this paper, we define a new task called zero-shot domain-specific question generation. Research on this problem helps to quickly build a dialogue system in specific application domains. To tackle this problem, we propose a novel model ZQNet to generate questions from triples, of which the predicates are unseen in the training dataset. We train ZQNet model on open-domain dataset and perform the zero-shot fine-tuning. Given structured data in a specific domain, the trained model can automatically generate reasonable descriptions. In order to evaluate the performance of different models, we construct a testing dataset that contains factual triples and corresponding questions in traffic domain and conduct a series of comparative experiments on the testing dataset, which shows that our model achieves reasonable performances and outperforms other baselines.