Explicitly Integrating Judgment Prediction with Legal Document Retrieval: A Law-Guided Generative Approach

TL;DR

The paper tackles the need for explicit judgment-consistent legal document retrieval. It introduces GEAR, a law-guided generative retrieval framework that explicitly integrates judgment prediction by extracting rationales, constructing a law-structure constraint tree, and employing a revision loss to align predictions with labels. GEAR generates law-aware semantic IDs and traverses the legal hierarchy to jointly predict relevant documents and applicable judgments in a single inference. Experiments on Chinese LCR datasets and a French SAR dataset show state-of-the-art retrieval performance with competitive judgment prediction and improved transparency and efficiency in the decision-making process.

Abstract

Legal document retrieval and judgment prediction are crucial tasks in intelligent legal systems. In practice, determining whether two documents share the same judgments is essential for establishing their relevance in legal retrieval. However, existing legal retrieval studies either ignore the vital role of judgment prediction or rely on implicit training objectives, expecting a proper alignment of legal documents in vector space based on their judgments. Neither approach provides explicit evidence of judgment consistency for relevance modeling, leading to inaccuracies and a lack of transparency in retrieval. To address this issue, we propose a law-guided method, namely GEAR, within the generative retrieval framework. GEAR explicitly integrates judgment prediction with legal document retrieval in a sequence-to-sequence manner. Experiments on two Chinese legal case retrieval datasets show the superiority of GEAR over state-of-the-art methods while maintaining competitive judgment prediction performance. Moreover, we validate its robustness across languages and domains on a French statutory article retrieval dataset.

Figures (5)

• Figure 1: The overview of our proposed GEAR. It mainly consists of three modules, including rationale extraction, law structure constraint tree, and revision loss. The middle part in blue is the generative retrieval framework.
• Figure 2: The proposed three modules of GEAR. $f_R$ extract rationales from legal documents; $f_T$ assigns hierarchical IDs to each document and constrain the decoding; $\mathcal{L}_c$ jointly optimizes document retrieval and judgement prediction.
• Figure 3: Comparison of retrieval performance of DSI and DSI equipped with the proposed three modules. The means values of 5 repeated experiments are reported, with error bars representing the 95% confidence interval of the means.
• Figure 4: Comparison of the judgment prediction performances in terms of coverage. The proposed GEAR consistently outperforms DSI and DSI-QG in both single (ELAM) and multiple (LeCaRDv2) charges scenarios.
• Figure 5: Testing curves of DSI, DSI-QG, and our GEAR.