Uncertainty-aware evidential learning for legal case retrieval with noisy correspondence

Abstract

Legal case retrieval is a critical task in intelligent legal systems, providing relevant precedents to assist judges in their decision-making. While current data-driven neural retrieval methods have demonstrated impressive performance on clean, annotated data, they often ignore the robustness against noisy correspondences. In practice, legal annotators are required to identify legal uncertainty, which refers to the ambiguity or unpredictability in legal interpretations and applications, in relevance estimation between cases. This uncertainty often introduces noise into the training data, leading to unreliable predictions and potentially impacting the fairness and justice of downstream tasks. Focusing on this robustness issue, we propose a novel evidential learning framework called ELCR, which explicitly models the legal uncertainty and addresses noisy correspondences. Specifically, we first estimate the multi-faceted relevance between query-candidate cases from the concept, rule, and fact levels. These relevance estimations are then used to obtain the evidence-based uncertainty under the Dempster-Shafer Evidence Theory, which helps correct labels from noisy correspondence. Guided by two elaborate evidence-based training objectives, ELCR provides accurate uncertainty estimation, enhancing reliability and robustness. Extensive experiments on various noise proportions across two benchmark datasets demonstrate that our method exhibits robustness against noisy correspondences while maintaining competitive retrieval performance.