TAGNet: a tiny answer-guided network for conversational question generation

Conversational Question Generation (CQG) aims to generate conversational questions with the given passage and conversation history. Previous work of CQG presumes a contiguous span as the answer and generates a question targeting it. However, this limits the application scenarios because answers in practical conversations are usually abstractive free-form text instead of extractive spans. In addition, most state-of-the-art CQG systems are based on pretrained language models consisting of hundreds of millions of parameters, bringing challenges to real-life applications due to latency and capacity constraints. To elegantly address these problems, in this work, we introduce the Tiny Answer-Guided Network (TAGNet) based on the lightweight module (Bi-LSTM) for CQG. We explicitly take the target answers as input, which interacts with the passages and conversation history in the encoder and guides the question generation through the gated attention mechanism in the decoder. Besides, we distill the knowledge from larger pretrained language models into our smaller network to make the trade-off between performance and efficiency. Experimental results show that our TAGNet achieves a comparable performance with large pretrained language models (retaining 95.9 % of teacher performance) while using 5.7 × fewer parameters and 10.4 × faster inference latency. TAGNet outperforms the previous best-performing model with similar parameter size by a large margin, and further analysis shows that TAGNet generates more answer-specific conversational questions.