Support/Attack Relation Classification

The foundational taxonomy of argument relations determines model architecture. Binary classification distinguishes support from attack, while multi-class schemas add nuance:

• Support: Premise strengthens conclusion
• Attack: Premise undermines conclusion (rebuttal, undercutter, or undermining)
• Neutral: No logical bearing on the target

Legal corpora often require fine-grained labels to capture the rhetorical complexity of judicial reasoning, where a citation may partially support while distinguishing a key point.

Support/attack classification typically operates as the final stage in a multi-step NLP pipeline:

1. Argument Component Detection: Identify premises and conclusions
2. Argument Pairing: Generate candidate relation pairs
3. Relation Classification: Predict the rhetorical link type

Modern end-to-end transformer models can jointly perform component detection and relation classification, reducing error propagation that plagues pipelined architectures in legal domains.

Effective classification depends on how argument pairs are encoded for the model:

• Concatenation encoding: Source and target arguments are joined with separator tokens
• Siamese encoding: Each argument is encoded independently before combining representations
• Graph-aware encoding: Argument position within the broader argument graph informs the relation

Legal texts benefit from document-level context windows that capture the full reasoning chain, not just adjacent sentence pairs.

Legal discourse contains explicit discourse markers that signal argumentative relations with high reliability:

• Support indicators: "therefore," "consequently," "as established by"
• Attack indicators: "however," "on the contrary," "the appellant erroneously contends"
• Concession patterns: "while it is true that... nevertheless"

Domain-specific lexicons for common law reasoning significantly outperform general-purpose discourse markers, as legal writing employs unique rhetorical conventions.

Legal argumentation frequently spans multiple filings, requiring cross-document classification:

• A complaint's claim is attacked by a motion to dismiss
• An appellate brief supports its argument with a trial court's finding
• An amicus brief provides neutral context to a constitutional question

This demands document-level embeddings and coreference resolution to track arguments across docket entries, not just within a single opinion.

Standard metrics for support/attack classification include:

• Macro F1: Handles class imbalance common in legal corpora where support relations dominate
• Directed accuracy: Measures whether the model correctly identifies the source→target direction
• Transitive consistency: Evaluates whether predicted relations form logically coherent argument graphs

Key legal benchmarks include the ECHO and Mochi datasets, which annotate U.S. Supreme Court and European Court of Human Rights decisions with argumentative structure.