Abstract
Rhetorical Role Labeling (RRL) is essential for legal document analysis but is hindered by the complexity and variability of legal texts. We present HiCuLR, a hierarchical curriculum learning framework that structures training by gradually increasing the difficulty of both input documents and target labels. By nesting a document-level curriculum (DC) within a rhetorical role-level curriculum (RC), HiCuLR allows models to learn fundamental structures before tackling fine-grained nuances. Evaluation across four Indian legal datasets demonstrates that HiCuLR consistently outperforms standard baselines and single-curriculum approaches.
Introduction
In Legal NLP, Rhetorical Role Labeling (RRL) involves identifying the functional role of each sentence in a judgment—such as "facts," "evidence," "reasoning," or "ruling." This task is a prerequisite for downstream applications like automated summarization and legal search.
Traditional training methods feed models examples in a random order, which can be problematic given the extreme variance in legal document structure and the subtle distinctions between certain labels. Complex, non-standard documents can introduce "noise" early in training, leading to sub-optimal local minima. HiCuLR addresses this by mimicking human cognition, providing models with a structured "syllabus" that starts with simpler concepts.
Methodology
HiCuLR implements two nested curriculum strategies to guide the learning process:
- Document-level Curriculum (DC): This orders input documents from "easy" to "hard." Difficulty is determined by metrics such as deviation from standard discourse or the statistical likelihood of label sequences. A "baby-step" scheduler gradually introduces more complex documents into the training pool.
- Rhetorical Role-level Curriculum (RC): This focuses on label difficulty. Early in training, the model uses "soft" labels where confusing similar roles (e.g., "analysis" vs. "ruling") is penalized less severely. These transition to "hard" one-hot encodings as training progresses, forcing the model to learn fine-grained distinctions.
The hierarchical nature of HiCuLR involves running a full cycle of the DC for every step of the RC, ensuring the model masters document structure at each level of label granularity.
Experimental Results
The framework was evaluated on four datasets of Indian legal judgments: Build, Paheli, M-CL, and M-IT. As shown in the performance table below, the hierarchical approach consistently yields the best performance across all metrics.
Key findings include:
- Significant Gains: On the "Build" dataset, HiCuLR reached a Macro-F1 of 63.61, compared to the 60.02 baseline.
- Synergy: The combination of RC and DC (rows 7 and 8 in the table) outperformed either strategy used in isolation, confirming that both input and output ordering are critical.
- Data-Driven Metrics: Statistical difficulty metrics, such as "Data-based Probabilistic Discourse," often proved more effective than expert-defined templates. This suggests that models find different patterns "easy" than human experts might predict.
Discussion and Conclusion
HiCuLR demonstrates that structured learning significantly improves model performance on complex legal tasks. By excluding the negative impact of difficult examples in early stages, the model finds better local minima in the parameter space.
However, limitations remain. Currently, the model assigns exactly one label per sentence, whereas complex legal sentences may serve multiple roles. Furthermore, the study focused on Indian legal documents; future work will explore the generalizability of these curricula to other jurisdictions (e.g., US or UK) and investigate sub-sentence level segmentation.
Summary updated