Toward Deployment-Oriented Long Legal Document Classification: A Segmentation-Based Framework for Distributed Legal Evidence Integration

Long-sequence models can improve long legal document classification. However, their high computational cost limits practical deployment in privacy-sensitive legal environments that favor on-premise solutions. To address this constraint, this paper proposes a segmentation-based framework. The framework uses Legal-BERT for segment encoding and incorporates segment overlap, together with an aggregation module composed of BiGRU modeling and attention-based aggregation. Experiments on the ECHR dataset under a leakage-controlled setting show that the proposed framework achieves 89.4% precision, 82.4% recall, and 84.6% F1-score, all macro-averaged. The results further suggest that the value of aggregation depends on how legal evidence is distributed across segments. When relevant evidence can be captured within individual segments, local prediction tends to favor precision. In contrast, when evidence is distributed across segments, BiGRU-based cross-segment modeling becomes important for improving recall. Attention-based aggregation alone is insufficient in this setting and becomes effective only after cross-segment dependencies are modeled. Additional results show that segment overlap improves performance, although excessive overlap provides no additional benefit. These findings highlight a practical trade-off in long legal document classification, where model effectiveness depends not only on architecture design but also on how segmentation and aggregation strategies support the recovery of distributed legal evidence.