Salience Scoring

Algorithms like LexRank model a document as a graph where nodes are sentences and edges represent semantic similarity (often cosine similarity of TF-IDF or embedding vectors).

• Eigenvector Centrality: A sentence is salient if it is similar to many other salient sentences.
• This method naturally identifies the most representative statements in a legal opinion without requiring prior domain knowledge.
• It excels at extractive summarization by selecting sentences that form the thematic hub of the document.

The foundational approach, TF-IDF (Term Frequency-Inverse Document Frequency), scores words based on how often they appear in a specific document versus a background corpus.

• A sentence's salience is the aggregate weight of its constituent words.
• In legal texts, this effectively surfaces terms of art like res ipsa loquitur or force majeure that define the document's core subject.
• While simple, it is highly effective for keyword-heavy documents like patents and contracts.

Modern models like BERTScore move beyond exact keyword matching by using deep contextual embeddings.

• A sentence's salience is measured by its cosine similarity to a vector representing the document's overall topic or a user's query.
• This captures paraphrased legal concepts, ensuring that a sentence stating 'the court affirmed the lower ruling' is recognized as salient to a query about 'appellate decisions'.
• Crucial for cross-document alignment where the same legal principle is expressed with different phrasing.

Legal documents follow strict rhetorical structures that provide strong prior signals for salience.

• Lead Bias: The first sentences of a judicial opinion or contract clause often contain the core holding or primary obligation.
• Segment Weighting: Text in defined sections like a contract's 'Definitions' or a brief's 'Argument Summary' is algorithmically upweighted.
• These heuristics are combined with statistical methods to ensure a document's structural logic is respected.

Salience is not always absolute; it is often dynamic and dependent on a specific information need.

• Maximum Marginal Relevance (MMR) scores sentences by balancing their relevance to a query against their redundancy with already-selected sentences.
• In legal research, this allows a system to build a summary that answers 'What is the standard for summary judgment?' by iteratively selecting the most relevant yet novel passages from a 50-page opinion.

When labeled data is available, salience scoring becomes a binary classification task.

• A model is trained on a corpus of legal documents where human experts have annotated each sentence as 'salient' or 'non-salient'.
• Features include sentence position, length, entity density, and semantic embeddings.
• This approach directly learns the nuanced annotation patterns of domain experts, often outperforming unsupervised methods for specialized tasks like ratio decidendi extraction.

A graph-based extractive summarization algorithm that computes sentence importance using eigenvector centrality. Sentences are nodes in a similarity graph; edges represent TF-IDF cosine similarity. The algorithm iteratively scores sentences based on how many other important sentences 'recommend' them, making it highly effective for identifying central legal holdings in multi-document case law analysis.

A query-focused summarization method that selects passages by balancing two competing objectives:

• Relevance: How well a passage answers the user's query
• Novelty: How different a passage is from already-selected content This redundancy penalty is critical in legal contexts where multiple documents may contain near-identical boilerplate language that should not dominate a summary.

An iterative prompting technique that generates increasingly dense summaries without increasing length. Starting from a sparse initial summary, each iteration identifies and fuses missing salient entities into the existing text. This produces entity-rich summaries ideal for legal professionals who need to rapidly identify all parties, statutes, and key dates mentioned across a document corpus.

The NLP task of identifying all linguistic expressions that refer to the same real-world entity. In legal texts, a single party may be referenced as 'the plaintiff,' 'Mr. Smith,' 'the appellant,' and 'he' across paragraphs. Accurate salience scoring depends on resolving these references to correctly aggregate importance signals for each entity rather than treating each mention as a distinct, low-importance token.

Recall-Oriented Understudy for Gisting Evaluation is the standard metric set for evaluating summary quality by counting overlapping n-grams between a candidate summary and a human-written reference. Variants include:

• ROUGE-1: Unigram overlap
• ROUGE-2: Bigram overlap (captures phrasing)
• ROUGE-L: Longest common subsequence (captures sentence structure) Salience scoring models are often tuned to maximize ROUGE scores against attorney-written summaries.

The degree to which a generated summary accurately reflects the stated facts of the source document without contradiction or fabrication. In legal AI, high salience scores are meaningless if the selected content is misrepresented. Factual consistency is verified using Natural Language Inference (NLI) models that check whether each summary claim is entailed by the source text, ensuring that salient extraction does not introduce hallucinated holdings.