Notes on MRS and the ERG

I've been reading through the literature on linguistic computation. This is a short review of Minimal Recursion Semantics and the English Resource Grammar: what they are, how they fit together, and what I've understood so far.

What MRS is

Minimal Recursion Semantics is a framework for representing sentence meaning in a form suitable for parsing and generation. The authors describe it as a meta-level language for describing semantic structures in an underlying object language. In the introductory paper, that object language is predicate calculus with generalized quantifiers ^[1].

The representation is flat. Meaning is a bag of elementary predications, where each predication is a single relation with its arguments. The predications are not nested inside one another. Scope relationships are reified with handles, which act as labels and as slots that can be filled later ^[1].

That flatness is the important design choice. A parser can produce a semantic structure without deciding every scope question up front. The scope information is represented, but it can remain unresolved until something downstream needs it.

MRS is the representation framework, not a parser or a reasoning engine. It defines the shape of the semantic structure that a grammar or processing tool can produce.

A small example

Here is a simplified example in MRS notation for "Every student read a book." It is written to show the shape of the representation rather than the exact output of a parser:

h0: every(x, h1, h2)
h1: student(x)
h2: some(y, h3, h4)
h3: book(y)
h4: read(x, y)

Each line is an elementary predication. The representation designates h0 as the top handle. The name is a convention, not something the handle carries intrinsically.

The quantifier every takes a restriction handle h1 and a body handle h2. The quantifier some, standing in for the existential contribution of "a," takes a restriction handle h3 and a body handle h4.

The body handles are what make the scope explicit. Here, h2 is filled by the some predication and h4 is filled by read. That gives the first reading.

Reading one, every outscopes some: Each student read some book or other. Different students may have read different books.

Reading two, some outscopes every: There is one particular book that every student read.

To represent both readings at once, an underspecified MRS would replace the relevant body handles with distinct holes and add constraints on how those holes may be filled. It is not simply the same graph with some lines erased. Resolving the constraints recovers the possible readings.

A real parse would also use predicate names such as _dog_n_1, event variables for verbs, handle constraints such as qeq, and variable properties such as tense. This example omits those details to keep the structure visible.

What MRS leaves to other layers

MRS is not itself a semantic theory. The introductory paper is explicit about this: MRS is a representation language that relates to an object language ^[1]. That distinction matters.

MRS represents semantic content. Drawing conclusions from that content requires additional interpretation and reasoning machinery. The introductory paper takes predicate calculus as its object language, and the translation from MRS into that language is a separate step from reasoning over the result ^[1][3].

MRS is also not a discourse representation or a knowledge base. It represents the compositional meaning of a phrase or sentence. Truth maintenance, belief revision, and discourse state are outside its scope.

None of that makes MRS a bad formalism. It means the representation is one layer, not the whole system.

What the ERG is

The English Resource Grammar is the large-scale grammar built around MRS. It is a broad-coverage HPSG grammar for English whose semantic representations are expressed in MRS ^[2].

It is not a syntax component followed by a separate semantics module. The grammar constrains syntax and semantics together, and processing tools such as ACE and the LKB use it to construct analyses that include both ^[2].

The ERG includes a hand-built lexicon, phrase structure rules, a type hierarchy, and a semantics interface designed for parsing and generation.

The grammar's semantic interface has been documented and regression-tested through a catalogue of compositional semantics ^[4]. That documentation is evidence of engineering discipline, and a reminder that precision requires sustained maintenance.

How they fit together

The ERG describes how English expressions are formed and how their meanings compose. A parser using the ERG produces a syntactic analysis together with an MRS representation. MRS supplies the representation language for the meaning side.

The grammar supplies linguistic precision. MRS supplies a compact representation that can leave some semantic decisions unresolved. The parser is what builds the analysis.

That combination is what makes the work interesting. The grammar and the representation are designed to work together, not bolted together after the fact.

What the work gets right

It is easy to focus on the limitations and conclude that the whole approach failed. That would be wrong.

MRS and the ERG established that compositional semantics can be implemented at scale. They established that underspecification is a practical engineering technique for parsing and generation. They established that a precise grammar can produce semantically rich structures that support transfer, generation, and analysis.

They also clarify the shape of the hard problems. Coverage is open-ended. Ambiguity has to be resolved somewhere. A representation is not a reasoner. Precision requires sustained engineering.

The useful lesson is not that MRS solved meaning. It is that representation and reasoning are separate jobs. The grammar builds the structure, MRS holds it, and something else has to reason over it. Keeping those layers distinct is more honest than pretending they are one thing.

References

1. Copestake, A., Flickinger, D., Pollard, C., and Sag, I. A. (2005). Minimal Recursion Semantics: An Introduction. Research on Language and Computation, 3, 281–332.
2. Flickinger, D. The English Resource Grammar. DELPH-IN documentation.
3. Fuchss, R., Koller, A., Niehren, J., and Thater, S. (2004). Minimal Recursion Semantics as Dominance Constraints: Translation, Evaluation, and Analysis. Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics.
4. Flickinger, D., Bender, E. M., and Oepen, S. (2014). Towards an Encyclopedia of Compositional Semantics: Documenting the Interface of the English Resource Grammar. Proceedings of the Ninth International Conference on Language Resources and Evaluation.