How linguistics learned to stop worrying and love the language models

Area of Science:

• Computational Linguistics
• Cognitive Science
• Artificial Intelligence

Background:

• Current debates question whether language models (LMs) truly learn language or offer insights into human linguistic processes.
• Some argue LMs negate the need for linguistic theory, while others doubt their relevance to human language learning.

Purpose of the Study:

• To evaluate the contribution of LMs to fundamental questions in linguistics, including structure, processing, and learning.
• To challenge extreme viewpoints on the capabilities and implications of LMs in linguistic research.
• To position LMs as valuable tools for exploring gradient, usage-based approaches to language.

Main Methods:

• Conceptual analysis of LM capabilities in relation to linguistic theory.
• Comparative examination of LM performance against established linguistic frameworks.
• Argumentation for LMs as model systems and proofs of concept.

Main Results:

• LMs can significantly contribute to understanding linguistic structure and human language processing.
• The success of LMs necessitates a re-evaluation of foundational concepts in linguistics.
• LMs provide empirical support for gradient, usage-based models of language.

Conclusions:

• Both extremes in the debate on LMs and linguistics are flawed.
• LMs do not replace linguistic theory but enhance it by serving as testable models.
• An optimistic perspective on the synergy between LMs and linguistics is warranted.