The question of how syntactic structure relates to semantic interpretation has occupied linguists since Montague first demonstrated that natural languages could be given the same rigorous semantic treatment as formal logical systems. The relationship between syntax and semantics—whether they constitute separate modules with a clean interface, deeply entangled systems, or something in between—remains unresolved. Recent work from both theoretical and computational perspectives is adding new dimensions to this longstanding debate.
The Research Landscape: Synthesis and Fragmentation
Monteza and Hermansyah (2025) provide a useful synthesis in their review paper, which draws together theoretical, empirical, computational, and cross-linguistic threads. Their central observation is that the interface question looks different depending on which linguistic tradition you start from: generativist approaches tend to treat syntax as primary, with semantics interpreting syntactic structures; cognitive-functional approaches see semantics as driving syntactic organization; and construction grammar blurs the boundary entirely.
The review highlights an underappreciated point: much of the disagreement about the interface stems not from empirical differences but from different definitions of what syntax and semantics are. If syntax is defined narrowly (phrase structure rules, movement operations), the interface appears relatively clean. If syntax is defined broadly (all structural regularities, including information structure and prosody), the boundary with semantics becomes diffuse.
Cross-Linguistic Evidence
Szabolcsi (2024) offers a selective but carefully argued set of cases where cross-linguistic data has been important to interface theory. Her examples include the role of Speaker and Addressee in grammar, mismatches between morphosyntactic form and semantic function, and the scopal behavior of quantifiers across languages.
A particularly instructive case involves quantifier scope. In English, "Every student read a book" is ambiguous: it can mean every student read the same book, or each read a different one. Many languages resolve this ambiguity syntactically—scope correlates with surface word order. But other languages (notably Hungarian, which Szabolcsi has studied extensively) show scope-word order mismatches that suggest the syntax-semantics mapping is not a simple surface-to-meaning correspondence. These cross-linguistic differences constrain which theories of the interface are viable: any adequate theory must accommodate both scope-transparent and scope-opaque languages.
The methodological implication is clear: the interface question cannot be settled by studying English alone. Data from typologically diverse languages—particularly those with free word order, rich morphology, or different scope-marking strategies—provides essential constraints.
The LLM Dimension
Kuczynski (2025) enters the debate from an unexpected angle, arguing that the success of large language models provides empirical support for classical theories of meaning, particularly the distinction between semantics and pragmatics. The argument goes roughly like this: LLMs achieve their linguistic competence by learning distributional patterns from text alone. If these distributional patterns are sufficient to approximate compositional semantic behavior, then something like compositional literal meaning must be a real property of language—not merely an artifact of formal theory.
This is a provocative claim, and it deserves careful scrutiny. The counterargument is straightforward: LLMs may achieve compositional-looking behavior through mechanisms that have nothing to do with compositionality as formal semanticists understand it. Statistical approximation of compositional outputs is not the same as compositional computation. Whether these different mechanisms matter—and for what purposes—is itself an open question.
Sattorova et al. (2025) provide a more grounded assessment of how computational linguistics has moved from rule-based syntactic analysis toward deeper semantic processing in the LLM era. Their observation is that while LLMs handle many syntax-related tasks well, they still struggle with tasks requiring genuine semantic understanding—negation scope, quantifier interactions, and metaphor processing among them. This pattern suggests that distributional learning captures some but not all aspects of the syntax-semantics mapping.
Critical Analysis: Claims and Evidence
<| Claim | Evidence | Verdict |
|---|---|---|
| Syntax and semantics are best understood as entangled rather than modular | Theoretical arguments + cross-linguistic scope data | ⚠️ Uncertain — depends heavily on how the modules are defined |
| Cross-linguistic data constrains interface theories | Szabolcsi's scope examples from Hungarian and other languages | ✅ Supported — different languages motivate different architectural assumptions |
| LLMs vindicate compositional semantics | Kuczynski's distributional learning argument | ⚠️ Uncertain — statistical approximation ≠ compositional computation |
| LLMs struggle with genuine semantic composition | Sattorova et al.'s task-based analysis | ✅ Supported — negation, quantifiers, metaphor remain challenging |
What the Disagreements Reveal
The disagreement between Kuczynski (who sees LLMs as evidence for classical compositional semantics) and the implicit conclusion from Sattorova et al. (whose findings suggest LLMs are limited in compositional semantics) is instructive. Both positions can be simultaneously correct: LLMs may vindicate the idea that compositional meaning is real (it leaves a distributional trace in text) while also demonstrating that distributional learning does not fully capture it (because composition requires more than pattern matching). The interface, in other words, may be real but not fully learnable from surface statistics alone.
Open Questions and Future Directions
What This Means for Your Research
For theoretical linguists, the message from this literature is that the interface question remains productively open. Neither radical modularity nor radical anti-modularity is well-supported; the interesting work lies in characterizing the specific ways structure and meaning interact.
For computational linguists, LLMs offer new tools for studying the interface—not as answers, but as probes. Their successes reveal which aspects of the mapping are distributionally recoverable; their failures reveal which aspects are not.
For typologists, this is a reminder that cross-linguistic data is not merely illustrative but constitutive of interface theory. The field needs more systematic typological work on scope, information structure, and the morphosyntax-semantics mapping.
Discover related work using ORAA ResearchBrain.