Extractive Summarization

The algorithmic backbone of extraction. Each sentence is assigned a numerical weight based on its importance to the central topic.

• TF-IDF Vectors: Scores words by frequency, penalizing common terms.
• Graph Centrality: Algorithms like LexRank treat sentences as nodes; centrality in the similarity graph determines importance.
• Positional Bias: Sentences near the beginning of legal documents often receive higher initial weights due to the structured nature of briefs and opinions.

A greedy selection algorithm that builds a summary by balancing two competing objectives.

• Relevance: How well a candidate sentence answers a query or matches the document's main theme.
• Novelty: How different the candidate is from sentences already selected.
• The Lambda Parameter: A tunable weight (0 to 1) that controls the trade-off. A high lambda favors relevance; a low lambda aggressively penalizes redundancy, which is critical for multi-document legal review.

A critical pre-processing step before salience scoring. This NLP task identifies all expressions that refer to the same entity.

• Pronoun Resolution: Mapping 'he,' 'she,' or 'it' back to the named party or object.
• Entity Linking: Connecting 'the plaintiff,' 'Acme Corp,' and 'the appellant' as a single node.
• Impact: Without this step, a key sentence starting with 'The Court held...' might be scored low because it lacks explicit entity keywords, causing the summarizer to miss the central holding.

The process of ensuring the final extracted summary does not contain repetitive information, a common failure mode in legal texts where judges often restate facts.

• Cosine Similarity Thresholds: Candidate sentences with high vector similarity to already-selected sentences are discarded.
• N-gram Overlap: A simpler, high-recall filter that checks for exact phrase repetition.
• Subsumption Detection: Identifies when one sentence fully contains the information of another, keeping only the more comprehensive statement.

The technique of explicitly linking each extracted sentence back to its precise location in the source document.

• Citation Anchors: Each summary sentence retains a pointer to the original paragraph, page, or line number.
• Provenance Tracking: Essential for legal workflows where an attorney must instantly verify the context of an extracted quote.
• Auditability: Unlike abstractive summaries, extractive outputs allow for deterministic, byte-for-byte verification against the source, eliminating hallucination risks.

A specialized mode where the summarizer selects sentences that specifically answer a user's natural language question rather than providing a generic overview.

• Relevance Modeling: Uses semantic similarity between the query embedding and sentence embeddings.
• Legal Application: A litigator can query 'What was the standard of review applied?' and receive only the extracted sentences discussing the standard of review, ignoring procedural history.
• Contrast with Generic Summarization: Generic extraction aims for overall coverage; query-focused extraction optimizes for precision and recall against a specific information need.

A graph-based extractive summarization algorithm that computes sentence importance using eigenvector centrality. It models a document as a similarity graph where nodes are sentences and edges are weighted by TF-IDF cosine similarity. The algorithm applies a stochastic PageRank-like process to identify the most central sentences, which are then extracted verbatim to form the summary. This method is particularly effective for multi-document legal summarization because it identifies sentences that are both salient and representative of the corpus's consensus view.

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

• Relevance: Maximizing similarity to the user's query or information need
• Novelty: Minimizing redundancy with sentences already selected

The algorithm uses a tunable lambda parameter to control the trade-off. In legal applications, MMR is used to construct summaries that answer specific research questions while avoiding repetitive extraction of the same legal principle stated across multiple citations.

The process of assigning a numerical weight to each sentence or passage based on its importance to the document's central topic. Features used for scoring include:

• Positional heuristics: Sentences in introductions and conclusions often carry more weight
• Term frequency statistics: TF-IDF or BM25 scores
• Entity density: Concentration of named legal entities (parties, courts, statutes)
• Cue phrases: Domain-specific signposts like 'the court held that' or 'we conclude'

These scores form the ranking basis for extractive selection.

The standard automatic evaluation metric suite for summarization. ROUGE measures the overlap of n-grams between a candidate summary and one or more human-written reference summaries. Key variants include:

• ROUGE-1: Unigram overlap
• ROUGE-2: Bigram overlap, capturing phrase-level fluency
• ROUGE-L: Longest common subsequence, capturing sentence structure

Because extractive summaries copy source text verbatim, they often achieve high ROUGE scores against references that also quote the source, making this metric particularly relevant for extractive system evaluation.

The technique of explicitly linking each factual statement in a summary back to its precise provenance in the source document. For extractive summaries, this is inherent—every sentence is a direct quotation with a known location. This property makes extractive methods highly valued in legal contexts where citation integrity is paramount. Systems can annotate extracted sentences with paragraph numbers, page lines, or document section identifiers, enabling attorneys to instantly verify the context of any summarized statement.

An iterative prompting technique that generates increasingly dense summaries without increasing overall length. Starting from an initial sparse summary, the model iteratively identifies and fuses salient entities and relationships. While primarily an abstractive method, the technique's focus on entity density and information compression provides a useful framework for evaluating whether an extractive summary has captured all critical legal entities—parties, dates, statutes, and holdings—within its verbatim selections.