CTG:A Controllable Text Generation Method based on the Joint Work of Language Model and Text Classifier

  • Xuyuan Liang
  • , Lihua Tian
  • , Chen Li
  • , Zhang Mandi

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

4 Scopus citations

Abstract

Text generation is an important research field in natural language processing. Today pre-trained models can generate text with high readability and fluency. However, the generated text content is usually not restricted, and the generated content cannot be controlled. Therefore, there is an urgent need for a controllable text generation method. Traditional controllable text generation methods usually have the disadvantages of high training cost and difficulty in adjustment. We propose a lightweight and easy-to-adjust method to update the language model through the discrimination results of the text classification model on the language model, thereby completing controllable text generation. And we have solved the problem of imbalance and quality degradation in text generation. Experiments show that the quality of the text generated by this method can also reach a higher level with strong subjectivity, can achieves the purpose of controlling text generation. And this method is easy to adjust, suitable for various scenarios.

Original languageEnglish
Title of host publication2021 IEEE 3rd International Conference on Communications, Information System and Computer Engineering, CISCE 2021
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages817-820
Number of pages4
ISBN (Electronic)9780738112152
DOIs
StatePublished - 14 May 2021
Event3rd IEEE International Conference on Communications, Information System and Computer Engineering, CISCE 2021 - Beijing, China
Duration: 14 May 202116 May 2021

Publication series

Name2021 IEEE 3rd International Conference on Communications, Information System and Computer Engineering, CISCE 2021

Conference

Conference3rd IEEE International Conference on Communications, Information System and Computer Engineering, CISCE 2021
Country/TerritoryChina
CityBeijing
Period14/05/2116/05/21

Keywords

  • KL divergence
  • Language model
  • Text classification
  • Text generation

Fingerprint

Dive into the research topics of 'CTG:A Controllable Text Generation Method based on the Joint Work of Language Model and Text Classifier'. Together they form a unique fingerprint.

Cite this