A Legal News Summarisation Model Based on RoBERTa, T5 and Dilated Gated CNN

Summarising legal news is crucial for monitoring judicial sentiments and managing emergencies, by distilling accurate, concise and comprehensive insights from extensive, complex legal texts to enhance the efficiency of managing public opinions. However, current methods are limited in handling long-form texts. To address this, we propose a novel two-phase sentence-level summarisation model. It first converts paragraphs into sentence vectors using RoBERTa and average pooling layer, then selects the more relevant sentences using a Dilated Gated CNN layer. These extracted sentences are treated as a corpus for the T5 PEGASUS model to generate refined summaries. Experiments show our model significantly outperforms baselines on ROUGE metrics by effectively capturing relevant information from lengthy, intricate legal opinions to produce high-quality summaries.