Maximum Marginal Relevance (MMR)

MMR operates on a greedy selection principle formalized as:

MMR = argmax[ λ * Sim(D_i, Q) - (1-λ) * max Sim(D_i, D_j) ]

• D_i: Candidate document or passage under consideration
• Q: The user's query or information need
• D_j: Documents already selected for the summary
• λ (lambda): A tunable parameter between 0 and 1

When λ=1, the algorithm behaves as a standard relevance-only ranker. When λ=0, it maximizes diversity by aggressively penalizing similarity to already-chosen items.

The λ parameter is the critical control knob that determines the relevance-diversity balance:

• High λ (0.7-0.9): Prioritizes query relevance. Useful when the information need is narrow and focused, such as finding a specific contract clause.
• Low λ (0.3-0.5): Prioritizes novelty and diversity. Essential for multi-document legal summarization where covering distinct precedents matters more than repeating the same point.
• Default practice: Most legal AI implementations start at λ=0.5 and tune based on ROUGE and BERTScore evaluation against attorney-written summaries.

In legal contexts, a lower λ prevents the summary from redundantly restating the same holding from multiple citations.

MMR builds the summary one passage at a time using an iterative greedy algorithm:

1. Compute relevance scores for all candidate passages against the query
2. Select the single most relevant passage as the first item
3. For each subsequent selection, compute the MMR score for every remaining candidate, penalizing similarity to all already-selected passages
4. Repeat until the desired summary length is reached

This greedy approach is not globally optimal but is computationally efficient—O(n²) in the number of candidates—making it practical for large legal document collections where exhaustive search would be infeasible.

The effectiveness of MMR depends heavily on the choice of similarity metric:

• TF-IDF Cosine Similarity: The original and still widely-used approach. Represents passages as sparse vectors weighted by term frequency-inverse document frequency.
• Dense Embedding Similarity: Modern legal AI systems use embeddings from models like Legal-BERT or Instructor-XL to capture semantic similarity beyond keyword overlap.
• Hybrid Approaches: Combining lexical (BM25) and semantic (dense vector) similarity scores often yields the best results for legal text, where precise terminology matters.

For cross-document alignment in multi-case analysis, embedding-based similarity is essential because different courts may use distinct phrasing for identical legal concepts.

MMR is particularly well-suited for synthesizing multiple legal documents:

• Case law synthesis: When summarizing 50+ precedents on a legal question, MMR ensures the output covers distinct holdings rather than repeating the most-cited case.
• Contract portfolio review: Analyzing hundreds of vendor agreements for a specific clause type benefits from MMR's diversity mechanism to surface unique variations.
• Regulatory change detection: MMR can select a diverse set of changed provisions across multiple amended statutes, preventing the summary from focusing exclusively on one section.

In practice, legal AI platforms combine MMR with source attribution so each selected passage retains a citation link to its originating document.

Despite its widespread adoption, MMR has known limitations:

• Greedy suboptimality: The incremental selection can miss globally better passage combinations
• Static λ: A single λ value applies uniformly, but optimal trade-offs may vary across different sections of a summary
• No semantic compression: MMR selects full passages verbatim; it cannot fuse information from two passages into a single, denser statement

Modern extensions address these gaps:

• Adaptive MMR: Varies λ dynamically based on the information content already covered
• Neural MMR: Replaces the hand-crafted formula with a learned scoring function trained on attorney-annotated summaries
• Chain-of-Density integration: Uses MMR for initial passage selection, then applies abstractive compression to fuse selected content

A technique that verbatim copies the most salient sentences from a source document to form a summary. Unlike abstractive methods, it does not generate new text.

• Mechanism: Ranks sentences by importance using features like term frequency or graph centrality
• MMR Connection: MMR is a classic extractive algorithm that re-ranks sentences to penalize redundancy
• Legal Use Case: Pulling key holdings directly from judicial opinions without risking paraphrasing errors

A technique that generates new, concise phrasing to capture the core meaning of source text, potentially rephrasing or paraphrasing the original content.

• Mechanism: Uses sequence-to-sequence models to produce novel sentences not found in the source
• Contrast with MMR: MMR is extractive; abstractive methods can fuse information across sentences
• Legal Risk: Requires rigorous factual consistency checks to prevent hallucination in generated text

The task of generating a summary that specifically answers a user's natural language question rather than providing a generic overview of the document.

• MMR's Native Domain: MMR was explicitly designed for query-focused settings, balancing relevance to the query against novelty
• Legal Application: Answering 'What is the standard of review applied?' from a 50-page appellate decision
• Key Distinction: Differs from generic summarization by incorporating a user information need into the salience model

A graph-based extractive summarization algorithm that computes sentence importance using eigenvector centrality in a similarity graph.

• Mechanism: Builds a graph where nodes are sentences and edges represent cosine similarity between TF-IDF vectors
• MMR Comparison: LexRank selects central sentences but does not explicitly penalize redundancy; MMR adds a diversity constraint
• Legal Use: Identifying the most representative sentences in a multi-document collection of related case law

The process of synthesizing information from multiple source documents into a single, coherent, and non-redundant summary.

• Core Challenge: Identifying and resolving overlapping or contradictory information across documents
• MMR's Role: MMR's redundancy penalty is critical for multi-document settings where the same fact may appear verbatim in several sources
• Legal Scenario: Merging factual backgrounds from a trial court opinion, appellate brief, and amicus filings into one timeline

A set of metrics that automatically evaluate summary quality by counting overlapping n-grams between a candidate summary and a human-written reference.

• Variants: ROUGE-1 (unigrams), ROUGE-2 (bigrams), ROUGE-L (longest common subsequence)
• MMR Evaluation: MMR-generated summaries are often benchmarked using ROUGE scores against gold-standard legal abstracts
• Limitation: ROUGE measures lexical overlap, not factual consistency or legal reasoning quality