Balancing Relevance and Diversity with Maximal Marginal Relevance

29 concepts25 papers2 proposals

In information retrieval, the "best" result is rarely a single item. If a user searches for an ambiguous term like "Jaguar," a list containing ten different pictures of the same car model is objectively less useful than a list featuring the car, the animal, and the professional football team. This tension between relevance (how well an item matches the query) and diversity (how much new information an item adds to the set) is the central challenge of result re-ranking.

The most enduring solution to this problem is Maximal Marginal Relevance (MMR). Rather than simply picking the top-$k$ most relevant items, MMR treats ranking as an iterative selection process that penalizes redundancy.

The Greedy Logic of Marginal Relevance

The intuition behind MMR is that the value of an item depends on what has already been seen. As Ah-Pine et al. describe, this is often implemented as a greedy selection algorithm. The process begins by picking the single most relevant item. For every subsequent slot, the algorithm calculates a "marginal relevance" score for all remaining candidates.

This score is a weighted combination of two factors:

  1. Relevance: The similarity of the candidate to the original query.
  2. Novelty: The dissimilarity of the candidate to the items already selected for the final list.

By tuning a parameter (often denoted as $\lambda$), developers can decide whether the system should prioritize pure precision or a broad variety of results. As Csurka et al. and Renders note in their work on social image retrieval, this simple iterative approach is remarkably robust, often outperforming complex clustering methods by providing a direct mechanism to penalize redundant information.

Reducing Redundancy in Summarization and QA

The need for diversity extends beyond search engines into natural language generation. In multi-document summarization, the goal is to condense vast amounts of text without repeating the same facts. Penn et al. highlight that relevance-diversity balancing is essential for speech summarization, where it prevents the final summary from being dominated by a single long-winded utterance.

Similarly, in complex question answering, Shima et al. demonstrate that applying MMR to candidate text segments ensures that the generated answer is concise and covers multiple facets of a question rather than circling a single data point. This principle is even applied to social media; Duan et al. use MMR to summarize Twitter topics, filtering out the "noise" of near-duplicate tweets to find a representative set that captures the evolution of an event.

The versatility of diversity-aware re-ranking is evident in its application to highly specialized fields. In virtual screening for drug discovery, Wang et al. use MMR to ensure that the "top" chemical compounds identified by a model are structurally diverse. This prevents researchers from wasting resources on minor variations of the same molecular scaffold, instead pointing them toward a variety of novel active compounds.

The mechanism even improves the underlying architecture of AI itself. Guo et al. applied marginal relevance to Approximate Nearest Neighbor (ANN) search. By treating neighbor selection as a diversification problem, they ensured that each edge in a search graph provides new "directional information," allowing search algorithms to explore the data landscape more efficiently.

Modern Adaptations: Neural MMR and RAG

As retrieval systems have moved toward deep learning, the MMR principle has evolved. Bi et al. introduced MMR-BERT, which adapts the principle into a neural framework for conversational search. Their system uses negative feedback—what the user didn't like—to pivot the search, penalizing candidates that are semantically similar to rejected items.

In the era of Large Language Models (LLMs), MMR serves as a critical component of Retrieval-Augmented Generation (RAG). For instance, Seo et al. utilize MMR when retrieving Korean legal statutes and precedents. By ensuring the retrieved context is diverse and non-redundant, the system provides the LLM with a broader factual base, which helps mitigate hallucinations and improves the accuracy of legal interpretations.

Why It Works: Information Coverage

The success of these various implementations, from APIRec's recommendation of niche web tools to ICTNET's multi-dimensional search diversification, stems from a shared goal: coverage.

As Haolun Wu et al. argue in their comprehensive survey, diversification bridges the gap between user-centric needs (avoiding "filter bubbles") and system-centric goals (fairness and discovery). By explicitly balancing what is relevant with what is new, MMR ensures that a limited number of result slots provide the maximum possible utility to the user.

Key Foundational Work

Evaluation of Diversity-Focused Strategies for Multimedia Retrieval (2008) — contributed 1 concept elements to this direction.
A Critical Reassessment of Evaluation Baselines for Speech Summarization (2008) — contributed 1 concept elements to this direction.
XRCE's Participation to ImageCLEF 2008 (2008) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_maximal_marginal_relevance_diversity_awa_7fe397ed,
  title = {maximal marginal relevance (diversity-aware reranking)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/maximal-marginal-relevance-diversity-aware-reranking/rdr_7fe397edbc844526b75a3dceea661e21},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_maximal_marginal_relevance_diversity_awa_7fe397ed_v1,
  title = {maximal marginal relevance (diversity-aware reranking)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/maximal-marginal-relevance-diversity-aware-reranking/rdr_7fe397edbc844526b75a3dceea661e21/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake