Legal Text Summarization using Bart and Explainable AI Techniques

Abstract

Summarizing lengthy documents, especially in the legal domain, posed significant challenges for humans and automated systems. Human efforts entailed considerable time and effort, while automated systems sometimes faltered in decision-making, leading to ambiguity in the generated summaries. This research explored the use of text summarization in legal documentation coupled with an explainability feature. It addressed the challenges of condensing lengthy legal texts and improving automated summarization systems' transparency. The research involved gathering legal documents, developing a Bidirectional and Auto-Regressive Transformers (BART) summarization model, and integrating explainability within the system, visualizing the attention mechanism. The system performance, which included BERT Score, cosine similarity, and Recall-Oriented Understudy for Gisting Evaluation (ROUGE) score between human-generated and system-generated summaries, and evaluation by target users, led to several engaging insights on legal summarization. The model demonstrated moderate performance, where user feedback indicated satisfaction with its functionality but highlighted the need for user interface improvements. Future improvements were suggested, including refining model training, enhancing the user interface, and adding features like adjustable summary lengths and language translation.