The Continuous Spectrum of Grammaticality

41 concepts28 papers1 proposal

In traditional formal linguistics, a sentence is often treated like a mathematical equation: it is either correct or incorrect, grammatical or ungrammatical. However, native speakers rarely experience language this way. We frequently encounter sentences that feel "marginal"—not quite wrong, but not entirely right either.

This research direction argues that linguistic acceptability is not a binary state, but a continuous gradient. By treating grammar as a system of "soft" probabilistic constraints rather than "hard" logical rules, researchers are building models that more closely mirror human linguistic intuition.

The Failure of Binary Categories

The core problem with binary grammaticality is that it cannot account for the nuances of human processing. As Owen Rambow observes in the study of the German "Third Construction," sentences can conform to formal rules yet still feel awkward or difficult for speakers to process. This gap suggests that "legal" word orders exist on a spectrum of acceptability.

Standard symbolic theories struggle with this nuance. To address this, Légendre et al. proposed Harmonic Grammar, a framework that treats grammatical rules as quantitative tendencies. In this view, a sentence’s "well-formedness" is the sum of satisfied and violated constraints, each weighted by its importance. This allows for a formal representation of why some errors feel minor while others render a sentence incomprehensible.

Quantifying Intuition

If acceptability is a gradient, how do we measure it? Traditional linguistics relied on the binary judgments of individual experts, but Munro et al. demonstrated that large-scale crowdsourcing provides a more accurate picture. By using fine-grained Likert scales or continuous sliders, researchers can capture a distribution of native speaker intuitions that reveals the probabilistic nature of linguistic knowledge.

Computational models have begun to adopt these continuous scales. For instance, Heilman et al. developed systems to predict grammaticality on a 1–4 ordinal scale, recognizing that distinguishing between a minor slip and total incomprehensibility is vital for tasks like automated essay scoring. Similarly, Cremers and Chemla used continuous slider scales to show that certain grammatical constraints—like those governing emotive-factive verbs—are "soft" degradations rather than absolute prohibitions.

Probabilistic Models as Proxies for Acceptability

One of the most powerful insights in this field is that the probability assigned to a sentence by a language model can serve as a proxy for its human acceptability. Lau et al. showed that by normalizing the output of unsupervised models (like RNNs), machines can achieve correlations with human judgments that rival the consistency of individual humans.

This works because language models naturally capture verb bias—the gradient preference humans have for specific syntactic structures based on the words involved. As Hawkins and Yamakoshi found, modern Transformers are particularly adept at capturing these subtle "vibes," reflecting how a specific verb might make one construction feel more natural than another, even if both are technically legal.

The Continuum of Meaning and Structure

The gradient approach extends beyond full sentences into the very building blocks of language:

  • Argumenthood: The distinction between a verb's essential "arguments" and optional "adjuncts" is often taught as a dichotomy. However, Kim et al. and Marchal and Poibeau argue it is a continuum. Some prepositional phrases are more "necessary" to a verb's meaning than others, and modeling this as a regression task improves the performance of semantic models.
  • Constructional Coercion: Verbs aren't static; they can be "coerced" into new meanings (e.g., "sneezing the foam off the coffee"). Bonial et al. show that the compatibility between a verb and a new construction is a gradient of probability, not an all-or-nothing switch.
  • Morphology: Even the "well-formedness" of new words (like computer-ism) is fuzzy. Hofmann et al. suggest that modeling the lexicon as a graph allows us to see how new words exist on a spectrum of plausibility.

The Influence of Context

Finally, acceptability is not a static property of a sentence in isolation. Bernardy et al. discovered a "compression effect" where the surrounding context makes poor sentences seem better and good sentences seem slightly worse. This suggests that human linguistic intuition is deeply influenced by pragmatics and memory, further complicating the attempt to draw a hard line between "correct" and "incorrect."

By moving away from binary labels, this research direction acknowledges that language is a living, probabilistic system. Whether in the evolution of historical dialects Seemann et al. or the pragmatic inferences of modern LLMs Pan et al., the shift toward gradience allows for a more precise and human-like understanding of how language actually works.

Key Foundational Work

Gapping and Frame Semantics: A fresh look from a cognitive perspective (1990) — contributed 2 concept elements to this direction.
Harmonic Grammar -- A Formal Multi-Level Connectionist Theory of Linguistic Well-Formedness: An Application ; CU-CS-464-90 (1990) — contributed 1 concept elements to this direction.
A Linguistic and Computational Analysis of the German "Third Construction" (1992) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_graded_grammaticality_46b1eecb,
  title = {graded grammaticality},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graded-grammaticality/rdr_46b1eecbc73f4999923a855281f94f1b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_graded_grammaticality_46b1eecb_v1,
  title = {graded grammaticality},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graded-grammaticality/rdr_46b1eecbc73f4999923a855281f94f1b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake