Argument Component Classification

The primary distinction is between Premises and Conclusions, but legal text demands finer granularity:

• Premise: A statement providing evidence or a reason to support a claim
• Major Premise: A statement of a legal rule, statute, or general principle
• Minor Premise: A statement of a specific fact or case detail
• Conclusion: The claim being advanced, often a proposed legal outcome
• Non-Argumentative: Text that provides background, procedural history, or citation without advancing reasoning

Unlike sentence-level rhetorical role labeling, this task operates at the sub-sentential token or phrase level. A single sentence may contain multiple components:

• "Because the defendant entered the property without consent, trespass has occurred."
• The first clause is a Minor Premise (factual assertion)
• The second clause is a Conclusion (legal claim)
• This fine-grained extraction enables precise reasoning chain reconstruction

Training data is created using rigorous argument annotation schemas that define:

• Token-level BIO tagging: Beginning, Inside, Outside tags for each component type
• Span boundaries: Precise rules for where a premise ends and a conclusion begins
• Inter-annotator agreement metrics: Typically requiring Cohen's Kappa > 0.8 for legal corpora
• Domain-specific guidelines: Rules for handling embedded citations, parentheticals, and multi-clause sentences unique to legal discourse

Component classification is the first stage in a multi-stage argument mining pipeline:

1. Component Classification: Identify and label spans (this task)
2. Relation Classification: Determine support/attack links between components
3. Argument Graph Construction: Build the full reasoning network
4. Reasoning Chain Reconstruction: Extract the inferential path to the final conclusion Errors at this stage propagate downstream, making high accuracy critical for reliable case outcome prediction and citation verification.

Modern systems employ domain-specific legal language models fine-tuned for sequence labeling:

• Transformer-based architectures: BERT variants pre-trained on legal corpora (CaseLaw-BERT, Legal-BERT)
• Conditional Random Fields (CRF): Often stacked on top of neural encoders to enforce valid label transitions
• Few-shot prompting: Using large language models with carefully engineered legal prompt engineering templates
• Data augmentation: Synthetic generation of annotated samples using synthetic data generation techniques for rare component types

Performance is measured using standard sequence labeling metrics adapted for legal argumentation:

• Token-level F1: Per-token accuracy of component boundaries and labels
• Span-level F1: Exact match of entire component spans, penalizing partial overlaps
• Macro-averaged F1: Equal weighting across all component types to prevent majority-class bias
• Key benchmarks include the ECHO and Mochi datasets derived from U.S. Supreme Court and European Court of Human Rights decisions