A Comparative Analysis of Language Generation Mechanisms in Cartesian Universal Grammar and Transformer-Based AI Models

Abstract

ABSTRACT

This paper provides a strict comparative point-of-view of two fundamentally opposed paradigms towards the perception and creation of human language: the nativist, computational-hierarchical, model of understanding human language as Cartesian Universal Grammar (UG) and operationalized by the Minimalist Program (MP), and the empiricist, statistical-associative, model of human language understanding as transformer-based Large Language Models (LLMs). The fundamental point of comparison is the driving generative processes of each model the structure-building process of the UG, which is Merge, and the mechanism of contextualization of the transformer, Self-Attention. Whereas UG assumes an innate domain- specific language faculty, which is discrete-infinity, recursively structure-generating, and characterized by discrete infinity, LLMs acquire their impressive language ability through statistical optimization on large corpora, using continuous, high-dimensional, vector representations.  To substantiate the dissimilarities between these two models, the analysis method applies a four-part framework, namely the Computational Primitive, Representational Structure, Source of Knowledge, and Explanatory Scope. Results have shown that UG has higher explanatory depth in terms of nature of language competence, and it gives a principled explanatory account on its formal properties, but on the other hand, LLMs exhibit higher predictive power and performance on language use, which is reflected in fluency, coherence, and scalability. Finally, the paper explains the implication of the empirical success of LLMs to the nativist hypothesis and suggests that a more detailed theory of language may well need a unification of the formal limitations that are inherent to UG with the statistical power of transformer architecture