Compositionality

Compositionality is the principle that the meaning of a complex expression is determined by the meanings of its constituent parts and the manner in which they are syntactically combined. It is the backbone of formal semantics: without compositionality, language would be an unstructured inventory of holistic signs, and systematic inference would be impossible. The principle states that semantics is homomorphic to syntax — the meaning of a whole is a function of the meanings of its parts and their mode of composition.

The principle is most commonly associated with Gottlob Frege, though the systematic formulation emerged in the twentieth century through the work of logicians and philosophers of language building on Frege's foundational insights. In formal semantics, compositionality is not merely a convenient assumption but a methodological necessity: it explains how finite minds can understand indefinitely many sentences, and it underwrites the possibility of truth-conditional semantics, where the truth conditions of a sentence are built up recursively from the referents of its terms and the logical properties of its connectives.

Compositionality and Structure

The power of compositionality lies in its economy. A language with compositionality can generate an infinite number of meaningful expressions from a finite vocabulary and a finite set of combinatorial rules. This is the same recursive power that makes syntax productive, and compositionality is the semantic counterpart: it ensures that the productivity of syntax is matched by a corresponding productivity of meaning. The principle connects the combinatorial structure of language to the inferential structure of thought: if 'all men are mortal' and 'Socrates is a man' compose to yield 'Socrates is mortal,' it is because the semantic composition mirrors the logical structure of the argument.

But compositionality is not merely a linguistic principle. It is a systems property. Any representational system that supports systematic generalization must have something like compositionality in its architecture. Neural networks that fail to exhibit compositional generalization — that can learn 'red square' and 'blue circle' but fail on 'red circle' — are systems that have not acquired the compositional structure of their training domain. The failure is not merely a performance limitation; it is an architectural diagnosis. The system has memorized correlations rather than learning the compositional rules that generate them. This connects compositionality to the broader theory of stochastic complexity, where the representational architecture of a system determines which patterns it finds simple and which it finds complex.

Limits and Challenges

Compositionality has limits, and the limits are informative. Natural language is full of apparently non-compositional constructions: idioms ('kick the bucket' does not mean what its parts suggest), metaphors ('time is a thief' composes syntactically but semantically fuses into something irreducible), and context-dependent expressions whose meaning varies with the speaker's intentions and the discourse situation. These phenomena have led some philosophers — notably proponents of contextualism — to argue that compositionality is at best an idealization and at worst a distortion of how language actually works.

The challenge, however, can be met without abandoning the principle. The apparent exceptions to compositionality often dissolve upon closer analysis. Idioms are compositional at a different level: they are single lexical items whose internal structure is frozen. Metaphors are compositional if we recognize that semantic composition operates over multiple dimensions — not just literal denotation but connotation, affect, and inferential potential. Context-dependence is not a failure of compositionality but an enrichment of it: the compositional machinery takes contextual variables as additional inputs, and the meaning of the whole is still a function of its parts and their mode of combination, just with a larger parameter space.

From a systems perspective, the debate over compositionality is a debate about the grain of analysis at which a system's behavior becomes predictable from its components. A system that appears non-compositional at one level may be perfectly compositional at another. The question is not whether language is compositional, but at what level of description compositionality obtains — and whether the levels at which it fails reveal something deeper about the nature of emergent semantics.

The compositionality wars in philosophy of language have been fought as if the principle were a binary property that language either has or lacks. This is a category error. Compositionality is not a fact about language but a methodological commitment — a choice to privilege a certain kind of explanatory architecture. Languages, like all complex systems, exhibit compositional structure at some levels and non-compositional dynamics at others. The error is not in asserting compositionality but in asserting it universally, as if the idioms and metaphors and contextual shadings that resist decomposition were mere noise rather than signals that the system is operating at multiple scales simultaneously. A theory of meaning that treats compositionality as a universal law is not a theory of language but a theory of formal systems that happen to look like language. The real work is to map the transitions between compositional and non-compositional regimes — and to understand what drives them.