Hierarchical Attention-Based GAN with Transformer for Abstractive Text Summarisation

The process of extracting important information from the original text source is also referred to as abstract-text summarisation (ATS). One long-term objective of artificial intelligence (AI) is to create an ATS system that can produce concise, relevant and credible summaries for the source texts. AI-based ATS frequently misses important details and fails to maintain the content’s original meaning in the generated summaries due to a lack of background knowledge. Current approaches also have trouble producing semantically rich summaries and accurately capturing hierarchical relationships in text. To overcome these challenges, this study proposes a novel approach for abstractive text summarisation using deep learning (DL) methods. Initially, the input documents are pre-processed using stemming, stop word removal, tokenisation and lemmatisation methods to remove unwanted data and increase the model’s flexibility. After pre-processing, multi-level feature extraction is carried out using the knowledge-based hierarchical attention (KBHA) module, which records semantic associations at the word, phrase and document levels. Furthermore, the suggested method makes use of a Laplacian generative adversarial network with Transformer (LGANsformer) to enhance the coherence and quality of text summarisation. LGANsformer combines the advantages of Transformer models with GANs to improve the quality of text summarisation. The approach achieves higher performance in capturing complicated textual structures and producing high-quality abstractions by merging KBHA and LGANsformer. The proposed approach is implemented in Python, and CNN/DailyMail, PubMed and Gigaword datasets are utilised to evaluate the results. The simulation results demonstrate that the proposed method yields superior results for abstractive text summarisation in terms of Rouge1, Rouge2 and RougeL.