Maximal Marginal Relevance (MMR)
Definition
Maximum Marginal Relevance (MMR) selects results stepwise to fit the query while minimizing redundancy.
MMR evaluates candidates, such as sentences, based on their relevance to a query and their similarity to already selected candidates. At each step, it chooses the candidate that maximizes this balance; a parameter λ weights relevance against redundancy. Cosine similarity is commonly used as the similarity measure in literature, but other measures are also possible in principle.
Simply explained
A reading list is compiled so that the texts fit the question and do not repeatedly state the same thing.
Why it matters in practice
Those who must select few from many hits can use MMR to limit repetitions while keeping the query in focus.
Example from business
Possible example: A search yields many similar hits; MMR selects a few that fit the query and do not nearly duplicate each other.
Limits and pitfalls
- The value of λ weights relevance against redundancy and must be chosen consciously. The similarity measure is interchangeable; cosine similarity is only one possible choice. MMR works stepwise (iteratively); Boudin et al. develop a variant that does not require iterative reordering.
Origin & history
Carbonell and Goldstein introduced MMR in 1998 at SIGIR. The work addressed document reordering and summary creation.
Use in mid-sized companies
- Possible application: A search team sets relevance and similarity measures as well as the MMR weighting. It evaluates result diversity and domain relevance on a fixed test set.
Frequently asked questions
Is MMR the same as ranking only by relevance?
No. MMR also considers how similar a candidate is to those already chosen.
What does λ stand for?
It is an interpolation coefficient between relevance and redundancy.
Is MMR used only for summaries?
No. The original paper's title also mentions document reordering.
Sources
- The use of MMR, diversity-based reranking for reordering documents and producing summaries · Carbonell & Goldstein (1998): The use of MMR, diversity-based reranking for reordering documents and producing summaries
- Improving Update Summarization by Revisiting the MMR Criterion · Boudin et al. (2010): Improving Update Summarization by Revisiting the MMR Criterion
Sources checked on 14.10.2026. Editorial application examples are hypothetical; no evidence of real customer results.
Editorially reviewed on: