Rhetorical Role Labeling

The task assigns a rhetorical role to every sentence in a judgment. Common labels include:

• Facts: The case's factual background and procedural history
• Arguments: Submissions made by the petitioner and respondent
• Statutes: References to relevant legal provisions
• Precedents: Discussion and analysis of prior case law
• Ratio Decidendi: The binding legal reasoning applied to the facts
• Final Decision: The verdict, sentencing, or relief granted

This granular segmentation enables precise information retrieval from lengthy documents.

Rhetorical Role Labeling is typically modeled as a token-level or sentence-level sequence classification problem. Modern approaches use:

• Hierarchical BiLSTM-CRF models that capture both local context and global label dependencies
• Transformer-based encoders (e.g., Legal-BERT, InLegalBERT) fine-tuned on annotated corpora
• Conditional Random Fields (CRF) as the final layer to enforce valid label transitions, preventing impossible sequences like a Final Decision appearing before Facts

Accurate rhetorical segmentation is a critical preprocessing step that unlocks higher-level legal AI tasks:

• Extractive Summarization: Automatically generate case briefs by selecting sentences tagged as Ratio Decidendi and Final Decision
• Citation Context Analysis: Isolate Precedent segments to understand how a court interpreted prior authority
• Argument Mining: Feed Argument segments into finer-grained claim and premise detection pipelines
• Case Outcome Prediction: Use the structured rhetorical flow as input features for predictive models

Rhetorical structures vary significantly across legal systems. A model trained on the Indian Supreme Court corpus (where the task originated) may underperform on US Federal or EU Court of Justice documents. Key challenges include:

• Structural divergence: Common law judgments follow different rhetorical conventions than civil law decisions
• Label set mismatch: Some jurisdictions require additional roles like Obiter Dictum or Procedural Directions
• Domain adaptation: Techniques like adversarial training and few-shot prompting are used to port models across jurisdictions with minimal re-annotation

The seminal dataset for this task contains 50 Indian Supreme Court judgments with ~12,000 sentences manually annotated by legal experts. Annotation follows a strict schema:

• Inter-annotator agreement is measured using Cohen's Kappa to ensure label consistency
• Contextual annotation: Annotators read the entire judgment before labeling to understand the document's rhetorical arc
• Sentence as atomic unit: Each sentence receives exactly one label, forcing a clean segmentation

This high-quality ground truth enables supervised learning with strong F1 scores exceeding 0.85 on held-out test sets.

Model performance is evaluated using standard sequence labeling metrics:

• Macro-F1 Score: Averages per-class F1 to account for label imbalance (e.g., Facts segments are often longer than Decision segments)
• Label Transition Accuracy: Measures whether the predicted sequence of roles follows a logically valid order
• Boundary Detection: Evaluates precision at segment transitions, critical for downstream segmentation tasks

State-of-the-art models achieve macro-F1 scores of 0.80–0.88 on the Indian Supreme Court benchmark, with Ratio Decidendi and Final Decision being the most reliably detected classes.