A Hybrid Transformer Framework For Multi-Label Classification Of Indian Legal Judgments

Abstract

The rapid growth of digitized judicial records in India has created a need for automatic methods that can organize judgments by legal domain. That requirement is difficult to satisfy because judicial decisions are long, semantically dense, and frequently span more than one area of law. Publicly accessible judgments also contain limited structured metadata, which reduces the usefulness of direct rule-based indexing. This paper presents a hybrid transformer framework for multi-label classification of Indian legal judgments. The proposed system combines LegalBERT-based contextual representations with stacked bidirectional long short-term memory layers so that both local semantics and document-level sequential cues can be captured from long judgments. A two-stage silver-standard annotation strategy is used to construct a scalable training corpus from Supreme Court judgments, while a manually curated gold-standard set is reserved for final evaluation. The framework further incorporates class-weighted binary cross-entropy, label-specific threshold calibration, and pseudo-label guided semi-supervised refinement to improve robustness under label imbalance and limited expert annotation. Experiments on ten legal domains show that the proposed model achieves stronger overall and minority-label performance than sparse-feature, embedding-based, and recurrent neural baselines. Ablation analysis also confirms that pseudo-label refinement, adaptive thresholding, and sequential modeling each contribute to the final gains. The study shows that a carefully designed hybrid architecture, supported by scalable annotation and imbalance-aware inference, provides a practical path for legal document classification in the Indian judicial domain.