Transformer-based Question Text Generation in the Learning System

Abstract

Question text generation from the triple in knowledge graph exists some challenges in learning system. One is the generated question text is difficult to be understood; the other is it considers few contexts. Therefore, this paper focuses on question text generation. Based on the traditional Bi-LSTM+Attention network model, we import Transformer model into question generation to get the simple question with some triples. In addition, this paper proposes a method to get the diverse expressions of questions (a variety of expressions of a question), that is, to take advantage of the semantic similarity algorithm based on Bi-LSTM with the help of a question database constructed in advance. Finally, a corresponding comparison experiment is designed, and the experimental results demonstrated that the accuracy of question generation experiment based on the Transformer model is 8.36% higher than the traditional Bi-LSTM + Attention network model.