Using Multiple Encoders for Chinese Neural Question Generation from the Knowledge Base

Abstract

Question generation is an important task in the field of natural language processing and intelligent tutoring system. Previous work on Chinese question generation focused on the rule-based approach, which requires a large amount of human resource to develop the question generation rules. With the recent success of deep neural network in natural language processing, especially the encoder-decoder neural network framework in machine translation, this study explored the effectiveness of the encoder-decoder network in Chinese question generation, where a triple from the knowledge base as an input is encoded and a question as the output is decoded. More importantly, the traditional encoder-decoder network is extended to have multiple encoders that can capture more diverse features to represent the triple. The study results showed that the model with multiple encoders outperformed the traditional encoder-decoder neural network by 1.78 BLEU points.