Abstractive Document Summarisation using Generative Adversarial Networks

Abstract

The use of automatically generated summaries for long texts is commonly used in digital services. In this thesis, one method for such document summarisation is created by combining existing techniques for abstractive document summarization with LeakGAN – a successful approach at text generation using generative adversarial networks (GAN). The resulting model is tested on two different datasets originating from conventional newspapers and the world’s largest online community: Reddit. The datasets are examined and several important differences are highlighted. The evaluations show that the summaries generated by the model do not correlate with the corresponding documents. Possible reasons are discussed and several suggestions for future research are presented.