Cognitive Condensation Through Adversarial Learning and Ranking in Abstractive Text Summarization

In recent years, there has been a significant increase in the use of neural networks for Abstractive Text Summarization (ATS). Notably, Generative Adversarial Networks (GANs) have demonstrated remarkable efficacy in domains such as computer vision and image generation. Recently, while using GANs for language generation problems, some researchers have observed encouraging results. Existing GANs, however, restrict the discriminator from functioning solely as a binary classifier, thereby impeding its capacity to effectively learn tasks involving outputs with complex structures, such as natural conversational descriptions. To overcome that limitation, we introduce a method incorporating a sentence ranking mechanism into generative adversarial networks to produce high-quality summaries. It is also well known that training GANs for discrete data, such as text, is challenging, so the policy gradient approach is used to optimize the generator. The presented approach yields \(Rouge-1\) as 41.58 and \(Rouge-L\) as 38.82 scores on a popular CNN/DM corpora, which are higher than the scores reported by earlier approaches. The proposed approach is also examined qualitatively, demonstrating its superiority over previous studies.