Graph-Enhanced Question Generation with Question Type Prediction and Statement Recovery

Shichen Chen, Shumin Shi*, Weiye Jiang

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Question Generation (QG) is an essential task in natural language processing, which aims to generate grammatical questions for given sentences or paragraphs. This task focuses on understanding the structures and semantics of sentences. This paper proposes a graph-enhanced question generation model based on Variational Autoencoders (VAE). We construct a semantic graph by dependency parsing and then encode the graph by iteration of Gated Graph Neural Networks (GGNN). Then we get the graph-enhanced representation by fusing the sentence-level and graph-level representation to perform pre-training of statement recovery and joint training of question type prediction and generation. Experiments show promising results of our proposed model on the most extensively used Question Answering (QA) dataset SQuAD.

Original languageEnglish
Title of host publicationProceedings - 2023 7th International Conference on Control Engineering and Artificial Intelligence, CCEAI 2023
EditorsDan Zhang, Yong Yue
PublisherAssociation for Computing Machinery
Pages60-65
Number of pages6
ISBN (Electronic)9781450397513
DOIs
Publication statusPublished - 28 Jan 2023
Event7th International Conference on Control Engineering and Artificial Intelligence, CCEAI 2023 - Sanya, China
Duration: 28 Jan 202330 Jan 2023

Publication series

NameACM International Conference Proceeding Series

Conference

Conference7th International Conference on Control Engineering and Artificial Intelligence, CCEAI 2023
Country/TerritoryChina
CitySanya
Period28/01/2330/01/23

Keywords

  • Gated Graph Neural Network
  • Multi-Task
  • Question Generation
  • Variational Autoencoder

Fingerprint

Dive into the research topics of 'Graph-Enhanced Question Generation with Question Type Prediction and Statement Recovery'. Together they form a unique fingerprint.

Cite this