Why is it called minimal recursion?

Its semantic representation minimizes nested structure by using relatively flat elementary predications with links and scope constraints.

Is an unresolved representation incomplete in a harmful way?

Not necessarily. It can accurately record what is known while leaving a genuine ambiguity open; whether resolution is needed depends on the task.

Is MRS only useful for parsing?

No. Its design also supports generation and semantic transfer, although particular systems implement different capabilities.

AI Glossary / Natural language and speech / Intermediate

Minimal recursion semantics

Definition

Minimal Recursion Semantics, or MRS, is a framework for representing sentence meaning using elementary predications connected by variables and scope information. Its relatively flat structure supports computational parsing and generation while allowing some semantic choices, especially quantifier scope, to remain unresolved until they are needed or further evidence becomes available.

What it means

A conventional logical expression can nest one meaning-bearing expression inside another. MRS separates smaller semantic contributions and records how they connect, instead of requiring every contribution to be embedded in a fully nested structure immediately.
Flat does not mean unstructured or devoid of scope. MRS preserves relationships through arguments, labels, and constraints. These can describe the allowable ways in which a more fully specified interpretation can be assembled.
The name concerns the representation of semantic structure. It does not mean that human language lacks recursion, that all processing becomes nonrecursive, or that meaning is reduced to an unordered collection of independent keywords.

How it works

A grammar associates words and constructions with elementary predications, often abbreviated EPs. An EP contains a predicate and its arguments, such as an event or participant variable.
During composition, contributions combine into a representation whose shared variables connect participants and properties. Labels identify predications or groups, while handle-valued arguments and constraints express scope relationships.
Some scope choices can remain underspecified. Rather than committing prematurely to one interpretation, the representation constrains the alternatives that are compatible with the analysis.
When an application needs a specific reading, an appropriate resolution process can supply further scope decisions. If the task does not require that distinction, retaining underspecification can avoid unnecessary commitment. Well-formedness and the grammar's conventions still constrain what counts as an acceptable representation.

Three practical examples

A grammar-based parser represents a sentence with quantifiers while retaining more than one permissible scope reading.
A generation system uses structured semantic contributions as input when producing a sentence that expresses the intended content.
A semantic-transfer experiment maps source-language contributions to target-language ones without making every irrelevant nesting choice part of the transfer rule.

Think of it this way

Think of storing the pieces of a structured plan with connection rules, while postponing a few ordering decisions. The pieces are organized enough to be useful before every permitted order is selected.
The analogy has limits: semantic scope is not merely the order of words on a page. It concerns how operators such as quantifiers affect interpretation, and the constraints have a formal role beyond an informal planning checklist.

Worked example

Consider the original example “Every technician inspected a device.” Two relevant readings are: each technician inspected some device, potentially a different one; or there is one device that every technician inspected.
Use a toy world with technicians Ada and Ben, devices D1 and D2, and exactly two inspection facts: Ada inspected D1, and Ben inspected D2.
The first reading is true because both technicians inspected at least one device. The second is false because neither D1 nor D2 was inspected by both technicians. Thus the two scope readings cannot simply be treated as equivalent.
An underspecified MRS can retain appropriate quantifier contributions and scope constraints without selecting one reading immediately. This example demonstrates why scope matters; it is not a complete serialized MRS. A complete representation would additionally specify the grammar's predicates, labels, arguments, and handle constraints.

Illustrative case study

In a fictional project, Alder Interfaces processes equipment statements for a question-answering demonstration. Its designers initially interpret every indefinite object as one shared object, even when the sentence also contains a universal quantifier.
They add examples that distinguish shared-object readings from potentially different-object readings. The semantic representation preserves the ambiguity when context has not resolved it, and the application requests clarification only when answering depends on the distinction.
The lesson is that postponing a semantic choice can be more faithful than making an unsupported one. This scenario is invented and reports no measured MRS system result.

Common misconceptions and mistakes

Do not describe MRS as a bag of words. Predicates, variable links, and scope constraints are essential parts of the representation.
Do not assume that flat structure removes semantic scope. MRS was designed to retain scope information while avoiding unnecessary structural embedding.
Do not treat underspecification as permission for every imaginable reading. The representation imposes constraints.
Do not confuse MRS with a complete theory of human meaning. The original authors describe a framework for representing semantic structures.
Do not infer that all ambiguities disappear automatically. Contextual resolution and application decisions may still be necessary.

Business use

For a business considering grammar-based language processing, MRS can provide an explicit intermediate representation for inspection and transformation. Its value depends on the language coverage, grammar, and application's semantic requirements.
Evaluate whether the system preserves distinctions that affect answers or actions, including participant roles, negation, and quantifier scope. Inspect cases where several interpretations remain possible rather than measuring only whether a parser returns a structure.
The framework does not by itself provide a production application, verified world knowledge, or a guarantee of complete sentence coverage. Those depend on the surrounding system and its evidence.

Check your understanding

In the toy world, does “each technician inspected some device” imply that one device was inspected by both?

Answer: No. Ada inspected D1 and Ben inspected D2, so the first statement is true while the shared-device statement is false. Preserving the scope distinction prevents that invalid conclusion.

Related terms

Semantic parsing
Elementary predication
Quantifier scope
Underspecification
Surface realization
HPSG

Learn next

Draw the two readings of a sentence containing “every” and “a,” then construct a small world where they differ in truth value.
Next inspect a documented MRS example and identify EPs, shared variables, labels, and handle constraints. Learn the formal scope conventions before editing a serialized structure; a plausible-looking graph is not automatically a valid MRS.

Sources

Copestake and colleagues, Minimal Recursion Semantics: An Introduction, author-hosted original paper: https://www.cl.cam.ac.uk/~aac10/papers/mrs.pdf DELPH-IN, official tutorial on formalisms and MRS underspecification: https://delph-in.github.io/docs/howto/DelphinTutorial_Formalisms/ The technician/device truth-condition example is an original illustration, not a serialized grammar output. Independent review: original scope/EP formulation and official underspecification tutorial consulted.

Previous: Minimal mappings Next: Minimax