A Systematic Literature Review on Multimodal Abstractive Text Summarisation Using Transformer-Based and Attention Models

The easy access to technology and the COVID-19 outbreak has accelerated communication among users through various multimedia formats (text, image, audio, and video). This has resulted in the mass production of multimodal information and has created challenges among many users in selecting relevant content, which is deeply buried within vast multimodal data generated. To address this issue, multimodal summarisation has emerged, aiming to enhance user experience by generating representative visual and textual summaries. However, current methods often lack semantic coherence and grammatical accuracy due to the complexity of integrating multiple modalities. This paper conducts a systematic review of multimodal abstractive texts summarisation techniques using PRISMA framework, with a focus on transformed-based models including attention mechanisms. It analyses recent studies, theirs methods, datasets, and performance based on ROUGE metrics. The findings indicate that approaches such as hierarchical attention and BERT-based models improve summarisation quality but still faced challenges including data alignment, computational cost, and limited datasets diversity. ROUGE benchmarking reveals inconsistent results with How2 and Dailymail datasets frequently used. The study highlights the need better modality fusion, improve generalisation and dataset diversity in future research.