Legal judgments and their complex content can be lengthy and difficult to read. Therefore, it is desirable but also risky if there are no explanations associated, to automate the summarization with the models as black box, not allowing human to intervene or edit their generated summary. This paper presents an explainable abstractive legal judgment summarization framework, combining transformer-based attention analysis with SHapley Additive exPlanations (SHAP). The model summarizes the lengthy legal judgments concisely while providing token-level explanations that map generated summaries to salient portions of legal texts. The proposed model yields an abstractive summarizer for Indian legal judgments reporting with ROUGE-1 value of 51.6, that of ROUGE-2 30.8 and 50.1 ROUGE-L surpassing that of extractive baselines and a higher correspondence between attention and SHAP scores making the trained model transparent to understand and trusted for explanation.
 

Legal document summarization; Explainable artificial intelligence; Transformer models; Attention mechanisms; SHAP; Legal NLP; Abstractive summarization; Judicial decision support