Clause-Level Summary

A technique that verbatim copies the most salient sentences from a source clause to form a summary without generating new text. In legal contexts, this approach preserves original phrasing—critical for maintaining precise contractual language.

• Relies on salience scoring algorithms like LexRank or TextRank
• Zero risk of hallucination since no new text is generated
• May produce choppy summaries lacking narrative flow
• Often used as a baseline before deploying abstractive methods

Generates new, concise phrasing to capture a clause's core meaning, potentially rephrasing or paraphrasing the original text. This mirrors how a human lawyer would explain a clause to a client.

• Enables plain-English restatement of complex legalese
• Requires robust factual consistency verification
• Often implemented with fine-tuned models like Legal-Pegasus
• Chain-of-Density prompting can increase information density without lengthening output

The degree to which a generated clause summary accurately reflects the stated facts of the source without contradiction or fabrication. This is the primary quality gate for legal AI.

• Measured via Natural Language Inference (NLI) models that check entailment
• Atomic Fact Decomposition breaks summaries into minimal claims for individual verification
• High factual consistency is non-negotiable for contract review workflows
• Directly impacts hallucination rate metrics in production systems

The technique of explicitly linking each factual statement in a generated summary back to its precise location in the source clause. This creates an auditable chain of reasoning essential for legal professionals.

• Enables one-click verification against original contract text
• Supports human-in-the-loop review workflows where attorneys validate outputs
• Often implemented via span-level annotations in the source document
• Critical for building trust with law firm and in-house counsel users

The NLP task of identifying all linguistic expressions that refer to the same real-world entity within a clause. Essential for correctly merging facts about a specific party across pronouns, definite descriptions, and named mentions.

• Resolves 'it,' 'such party,' 'the aforementioned' to the actual entity
• Prevents fragmented or duplicated entity references in summaries
• Particularly challenging in legal text due to formal anaphoric conventions
• Foundational preprocessing step before clause summarization pipelines

Recall-Oriented Understudy for Gisting Evaluation—a set of metrics that automatically evaluate summary quality by counting overlapping n-grams between a candidate summary and a human-written reference.

• ROUGE-1 measures unigram overlap; ROUGE-L uses longest common subsequence
• Provides rapid, automated quality signals during model development
• Limited in capturing semantic equivalence or paraphrasing quality
• Often supplemented with BERTScore for semantic-level evaluation in legal domains