Abstract
This article argues that the systematic character of language can be adequately examined only through a metalanguage capable of reflecting on language from a higher-order standpoint. It first considers formal (mathematical) logic as a candidate metalanguage: although it can capture truth relations among propositions, its restricted semantic range prevents it from serving as a metalanguage for the full spectrum of thought. It then revisits Wittgenstein’s notion of language games as an alternative metalanguage that translates meaning into a set of socially rule-governed uses. While this approach offers an external and publicly inspectable point of reference for semantics, it remains limited to reflecting on the nature of rules and does not account for language’s thought-laden dimension. The paper next turns to large language models (LLM) as a contemporary, non-traditional metalanguage. LLM operationalize a grammar of signs grounded in correlational relevance—probabilistically optimized linkages among tokens—and evolutionary updating, whereby empirical Bayesian adjustment continuously recalibrates the credibility of these correlations. On this view, LLM can process relations at scale, yet the resulting sign system remains a mapping that fails to represent the wholeness of objects and the world; it therefore lacks genuine understanding. The essence of thought lies in “selecting” a particular “possible world”. At its core, thinking is an act of “choice”: it selects and commits to a certain possible world among alternatives. Against this background, the article proposes a speculative “grammar of thought” and associated metalanguage built on negation and verbs. Negation secures the essential function of thought as selection—the capacity to choose among possible worlds—while verbs provide the generative structure through which possible worlds are constructed. The author suggests that this framework could upgrade mere mapping into a morphism oriented toward holistic intelligibility, offering a methodological route toward a next-generation of AI “world model”. Its putative advantage lies in simultaneously opening unlimited possibilities, supporting causal explanation, and enabling reflexive analysis of any language from within natural language itself, thus obtaining a better metalanguage.
Key words
metalanguage /
grammar of thought /
verb-centered thinking /
negation /
large language models
Cite this article
Download Citations
Zhao Tingyang.
A Philosophical Review on Metalanguage, the Grammar of Thought, and the “Grammar” of Large Language Models#br#[J]. Chinese Journal of Language Policy and Planning. 2026, 11(1): 5-15 https://doi.org/10.19689/j.cnki.cn10-1361/h.20260101
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}