Hierarchical Summarization

The first pass divides the source document into manageable segments that fit within the model's context window. Each chunk undergoes extractive or abstractive summarization independently.

• Chunk size is determined by the model's maximum token limit (e.g., 4,096 for older models, 128k+ for modern ones)
• Overlap windows between chunks prevent information loss at boundaries
• Each chunk summary captures the salient facts, parties, and legal principles within that segment
• This stage is embarrassingly parallel, allowing distributed processing of massive documents

The intermediate summaries from the first pass become the input for a second round of summarization. This process repeats until a final, coherent summary of the desired length is produced.

• A merge operation combines adjacent chunk summaries into larger context groups
• Each recursion level reduces total token count by a compression ratio (typically 3:1 to 10:1)
• The recursion tree can be depth-2 for most legal documents or depth-3+ for book-length texts
• This mirrors the map-reduce paradigm adapted for language model processing

Critical legal details can be lost across recursive steps. Retention mechanisms preserve key information that must survive compression.

• Entity extraction identifies and preserves named parties, dates, jurisdictions, and citations across all recursion levels
• Salience scoring assigns importance weights to facts, ensuring high-value information survives each compression pass
• Coreference resolution normalizes all references to the same entity (e.g., 'the plaintiff,' 'Mr. Smith,' 'he') before summarization
• A fact registry can be maintained as a sidecar data structure, tracking extracted assertions independently of the summary text

Each recursion level introduces risk of hallucination or distortion. Verification steps validate that intermediate summaries remain faithful to their source chunks.

• Natural Language Inference (NLI) models check whether each summary sentence is entailed by the source chunk
• Atomic fact decomposition breaks summaries into minimal claims for individual verification
• Source attribution links each claim back to its originating passage for human auditability
• A consistency threshold can trigger re-generation if the factual alignment score drops below an acceptable level

Legal documents present unique challenges that generic hierarchical summarization must address.

• Ratio decidendi extraction ensures the binding legal principle survives compression, while obiter dictum can be safely compressed
• Citation preservation maintains the integrity of legal references (e.g., 347 U.S. 483) across recursion levels
• Deontic logic markers — obligations ('shall'), prohibitions ('must not'), permissions ('may') — receive elevated salience scores
• Temporal sequences (effective dates, deadlines, statutes of limitations) are explicitly tracked to prevent chronological distortion

Hierarchical summarization quality is measured differently than single-pass approaches due to the compounding nature of errors.

• ROUGE and BERTScore measure lexical and semantic overlap with human-written reference summaries
• Hallucination rate is measured at each recursion level, not just the final output, to identify problematic compression stages
• Factual consistency scores compare each intermediate summary against its direct source chunk using NLI entailment
• Information retention curves track what percentage of original key facts survive each recursion level, identifying optimal compression ratios

A technique that verbatim copies the most salient sentences from a source document to form a summary without generating new text. In a hierarchical pipeline, extractive methods often serve as the first pass to identify high-value chunks before abstractive refinement.

• Selects sentences based on salience scoring and graph centrality
• Preserves original legal phrasing, reducing hallucination risk
• Algorithms include LexRank and TextRank
• Common in e-discovery and first-pass document review

A technique that generates new, concise phrasing to capture the core meaning of a source text, potentially rephrasing or paraphrasing the original content. In hierarchical architectures, abstractive models condense chunk-level summaries into a final coherent narrative.

• Relies on sequence-to-sequence Transformer models
• Enables cross-document fusion and redundancy elimination
• Requires rigorous factual consistency verification
• Prone to hallucination without grounding mechanisms

The degree to which a generated summary accurately reflects the stated facts of the source document without contradiction or fabrication. Natural Language Inference (NLI) models determine if a summary sentence is entailed by, contradicts, or is neutral to the source.

• Critical for legal admissibility of AI-generated summaries
• Measured via atomic fact decomposition against source text
• NLI serves as an automated fact-checking layer in pipelines
• Directly impacts hallucination rate metrics

Transformer models employing sparse attention mechanisms that scale linearly with sequence length, enabling processing of long legal documents without chunking. Longformer uses a sliding window plus global attention; BigBird combines random, window, and global attention patterns.

• Handle sequences up to 4,096 tokens (BigBird) and beyond
• Reduce quadratic memory cost of standard self-attention
• Serve as alternatives to hierarchical chunking for mid-length documents
• Often used as the base encoder in legal summarization systems

The technique of explicitly linking each factual statement in a generated summary back to its precise location in the source document. In hierarchical summarization, attribution chains must be preserved across multiple condensation layers to maintain auditability.

• Essential for citation integrity in legal AI
• Enables attorneys to verify claims without re-reading full documents
• Implemented via span-level provenance tracking
• Supports human-in-the-loop review workflows

An iterative prompting technique for generating increasingly dense and entity-rich summaries without increasing overall length. Each iteration identifies missing salient entities and fuses them into the summary, making it ideal for the final fusion stage of a hierarchical pipeline.

• Balances informativeness against conciseness
• Produces summaries suitable for headnote generation
• Entity-centric approach aligns with legal knowledge graph construction
• Works with both extractive and abstractive backends