Query-Focused Summarization

Unlike generic summarization which relies on global document statistics, query-focused systems dynamically recalculate salience scores based on semantic similarity to the user's query. A passage about 'venue' might be low-salience in a generic contract summary but becomes the highest-priority text when the query is 'Where must disputes be arbitrated?' This is typically implemented via cross-attention mechanisms between the query embedding and the source document tokens, or by using Maximum Marginal Relevance (MMR) to balance query relevance against redundancy.

A foundational algorithm for query-focused extractive summarization that greedily selects passages by optimizing a linear combination of two factors:

• Relevance: Cosine similarity between the candidate passage and the user's query
• Novelty: Penalty for similarity to already-selected passages (1 - max similarity to summary)

The λ parameter (typically 0.5-0.7) controls the trade-off. High λ favors relevance; low λ favors diversity. This prevents the summary from repeating the same information and ensures broad coverage of the query's aspects.

In multi-document settings, query-focused summarization must perform cross-document alignment to identify passages discussing the same entity or event across distinct sources. The system must then fuse this information, resolving contradictions and eliminating redundancy. For legal applications, this means synthesizing a coherent answer from multiple cases, statutes, and secondary sources—each with different precedential weight. Source attribution becomes critical here, linking each claim back to its origin document with precise citation.

A query-focused summary must remain factually consistent with source documents regardless of how the query reframes the topic. Modern systems employ Natural Language Inference (NLI) models as a post-hoc verification step: each atomic fact in the summary is treated as a hypothesis, and the source text serves as the premise. The NLI model classifies each fact as entailed, contradicted, or neutral. A high contradiction rate triggers regeneration. This is especially critical in legal contexts where a single hallucinated holding could misrepresent binding precedent.

An iterative prompting technique that produces increasingly information-dense summaries without increasing length. The process:

• Step 1: Generate an initial entity-sparse summary
• Step 2: Identify 1-3 missing salient entities not yet covered
• Step 3: Rewrite the summary to incorporate those entities while maintaining the same word count
• Repeat for 5 iterations This yields summaries with high entity density—critical for legal queries where missing a specific party, date, or clause reference renders the summary useless for downstream tasks.

A granular evaluation method where the generated summary is decomposed into minimal, self-contained factual claims—each a single subject-predicate-object triple. Each atomic fact is independently verified against the source document. For example, 'The court ruled that the patent was invalid due to prior art' decomposes into: (1) The court issued a ruling, (2) The ruling concerned a patent, (3) The patent was found invalid, (4) The grounds were prior art. This fine-grained verification catches partial hallucinations that sentence-level evaluation misses.

A foundational algorithm for query-focused summarization that selects passages by balancing two competing objectives: relevance to the user's query and novelty relative to already-selected content. The algorithm uses a tunable lambda parameter to control the trade-off.

• Relevance: Measured via cosine similarity between a candidate passage and the query
• Redundancy Penalty: Subtracts the maximum similarity between the candidate and any already-selected passage
• Greedy Selection: Iteratively builds the summary by picking the highest-scoring remaining passage

This prevents the common failure mode where a summary repeatedly restates the same fact while omitting other critical information.

The process of assigning a numerical weight to sentences, clauses, or passages based on their importance to the specific query rather than to the document as a whole. Unlike generic summarization, query-focused salience is context-dependent.

• TF-IDF Overlap: Measures lexical similarity between query terms and candidate text
• Semantic Embedding Similarity: Uses dense vector representations to capture conceptual relevance beyond exact keyword match
• Positional Bias: Legal documents often place key holdings in specific structural locations
• Entity-Centric Scoring: Boosts sentences mentioning parties, statutes, or legal concepts named in the query

Effective salience scoring is the primary differentiator between a generic summary and one that answers a specific legal question.

A critical post-generation step that validates whether every claim in a query-focused summary is entailed by the source document. This is especially vital in legal contexts where fabricated case citations or misstated holdings carry severe professional consequences.

• Natural Language Inference (NLI): Treats each summary sentence as a hypothesis and checks if the source text entails it
• Atomic Fact Decomposition: Breaks summary sentences into minimal claims for granular verification
• Source Attribution: Links each factual assertion back to its precise span in the original document
• Hallucination Rate: Quantifies the percentage of generated tokens that lack grounding in the source

Without this layer, query-focused systems risk producing highly relevant but legally false answers.

The task of identifying and linking semantically related passages across multiple legal documents that address the same query. This is essential when synthesizing an answer from a corpus of cases, statutes, and secondary sources.

• Entity Linking: Resolves mentions of the same legal entity, statute, or doctrine across documents
• Coreference Resolution: Identifies all expressions referring to the same real-world party or concept
• Temporal Alignment: Orders events across documents into a coherent chronological sequence
• Conflict Detection: Flags contradictory holdings or factual assertions between sources

Cross-document alignment transforms a collection of individual summaries into a coherent, multi-source synthesis that captures the full legal landscape relevant to the query.

An iterative prompting technique that generates increasingly information-dense summaries without increasing length. Starting from an initial sparse summary, each iteration identifies missing salient entities and fuses them into the text.

• Entity Identification: Locates key legal entities, doctrines, and facts not yet included
• Fusion Editing: Rewrites the summary to incorporate new entities while maintaining concision
• Density Constraint: Enforces a fixed token budget across iterations
• Legal Adaptation: Prioritizes inclusion of case names, statutory citations, and procedural postures

This technique is particularly effective for query-focused legal summarization where users need maximum informational yield from a constrained reading time.

A strategy for handling legal documents that exceed a model's context window by recursively summarizing chunks and then synthesizing those intermediate summaries. This enables query-focused summarization of entire multi-hundred-page contracts or appellate records.

• Chunk-Level Summarization: Each section is summarized independently with respect to the query
• Intermediate Fusion: Chunk summaries are merged into a coherent intermediate representation
• Final Synthesis: The merged intermediate summary is refined into the final query-focused output
• Context Retention: Critical cross-chunk dependencies are preserved through overlapping windows

This approach is essential for legal AI systems operating on full case files rather than isolated excerpts.