Cascading Circle Loss (CasCL)
This paper introduces Cascading Circle Loss (CasCL), a contrastive learning framework for legal retrieval tasks with graded relevance. CasCL combines an InfoNCE objective for positive-negative separation with a cascade of ranking boundaries that model ordered relevance levels between queries and legal clauses.
The framework also includes a negative-aware self-paced optimization mechanism and a relevance gain weighting strategy to improve ranking quality.
Experiments on legal passage retrieval benchmarks show that CasCL consistently outperforms existing methods, including InfoNCE, Circle Loss, MNRL and Margin MSE, across both binary retrieval and graded ranking metrics.
The effectiveness of CasCL stems from its ability to directly optimize for graded relevance, rather than just binary relevance.
This is achieved through the "cascade of ranking boundaries," which explicitly models the ordered nature of legal relevance judgments.
Furthermore, the inclusion of the negative-aware self-paced optimization helps the model focus on challenging negative examples, while the relevance gain weighting strategy ensures that improvements on highly relevant items are prioritized.
- Dataset: Legal Passage Retrieval (LPR)
Methods compared: InfoNCE, Circle Loss, MNRL, Margin MSE - Example Test Query - Statute Title: S.723 - Subject: Indian reservation; Contracts
- Result: CasCL achieved a 'LogLoss' of 0.372 and a 'MAP' of 0.305