Answers are grounded in this paper and cite their sources.
Legal Judgment Summarization aims to generate concise yet comprehensive summaries for complex and lengthy legal judgment documents. Current generation models often overlook low-frequency but critical legal details and suffer from memory decay over long documents. To address these challenges, we propose a novel insight: structural anchors of a legal judgment can capture its global key semantics, enabling models to retain sparse yet pivotal legal details while alleviating memory decay. Structural anchors constitute a distinct semantic representation beyond traditional bag-of-words or embeddings. To model the insight, we propose a Sequence-Structure Fusion (SSF) Framework, which leverages structural anchors for enhancing long-text sequence modeling. In the framework, to construct structural anchors, we design a Legal-guided Hierarchical Anchor Mining (LHA-Miner) model that leverages the cited statutory articles as weakly supervised signals to mine a multi-level taxonomy of anchors, constituting an extension of traditional topic modeling tailored to the legal domain. Second, to enhance long-text modeling, we introduce a Bidirectional Structural Semantic Propagation (BSSP-Integrator) strategy. This mechanism identifies related yet semantically distinct anchors through the structural anchors, diverging from existing methods focused solely on textual similarity. Finally, to generate high-quality summaries, we develop an Embedding-aligned Multi-Agent Collaborative Generator (EMAC-Gen), which explores multi-agent collaboration based on three defined quality criteria and utilizes anchor-augmented embedding signals as external supervision. Extensive experiments on two datasets show that the proposed SSF framework achieves significant metric improvements over all baselines.
No takes yet. Share an insight, caveat, or question.
Lyu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: