Style transfer for text simplification


Learning objectives: 

  • Gain a deep knowledge of different deep learning model architectures and style transfer methods for text simplification.
  • Explore and analyse the different deep learning model architectures, style transfer techniques and their adaptability for text simplification using non-parallel data. 
  • Gain advanced skills on using deep learning models for style transfer in text simplification. 
  • Acquire scientific writing skills through the preparation of the final thesis and presentations for the demonstration of the progress and results. 

Project description: 

Text style transfer (TST) is an important task in natural language processing focussing on changing the style specific properties in text like simplicity, politeness, and emotion while retaining the style independent content.  TST has a wide range of applications in natural language generation such as automatic text simplification, poetic writing , offensive language translation, and sentiment translation [1]  

With the success of deep learning in the last decade, a variety of neural methods have been recently proposed for TST and are heavily influenced by two related research areas: neural machine translation and neural style transfer (introduced in image processing). If parallel data, which includes aligned parallel sentences with the same meaning, but different styles are provided, standard sequence- to-sequence models which is the most commonly used neural machine translation approach, are often directly applied. However, most use cases do not have parallel data, so TST on non-parallel corpora has become a popular research area. Most of the TST approaches on non-parallel corpora aims to disentangle style and content attributes in text which has proven to be much harder than when it comes to images. Therefore, recently some methods have been proposed to perform TST without the disentanglement of content and style.  

Text Simplification aims to rewrite a complex sentence with simple structures while constrained by limited vocabulary. This task can be defined using TST, as reducing the expertise level from expert to layman language without modifying the content. Most of the initial systems for text simplification used neural machine translation considering the problem as a monolingual translation problem using parallel data. Recently, TST using neural style transfer has gained significant attention for text simplification. However, these methods also have their own challenges when applied to text simplification; limited availability of data, usability of the methods in text simplification, capability to disentangle style from content, etc.  

[1] Hu, Z., Lee, R.K.W., Aggarwal, C.C. and Zhang, A., 2020. Text Style Transfer: A Review and Experimental Evaluation. arXiv preprint arXiv:2010.12742. 


The scope of this honours project is to develop and comparatively assess algorithmic techniques for text style transfer. The primary focus will be on applications related to text simplification. The main outcomes of the project will be:

  • Literature review on text style transfer methods for non-parallel data. This includes understanding and reporting of the applicability of deep learning architectures, style transfer methods and challenges associated with them.
  • Development, evaluation and comparative assessment of state-of-the-art algorithms on selected datasets.
  • Demonstration on a selected style transfer problem such as medical text simplification.


  • Experience in python  programming
  • Knowledge on machine learning


Text simplification, style transfer, deep learning

Updated:  10 August 2021/Responsible Officer:  Dean, CECS/Page Contact:  CECS Marketing