How Linguistics Learned to Stop Worrying and Love the Language Models (2501.17047v2)

Abstract: LLMs can produce fluent, grammatical text. Nonetheless, some maintain that LLMs don't really learn language and also that, even if they did, that would not be informative for the study of human learning and processing. On the other side, there have been claims that the success of LMs obviates the need for studying linguistic theory and structure. We argue that both extremes are wrong. LMs can contribute to fundamental questions about linguistic structure, language processing, and learning. They force us to rethink arguments and ways of thinking that have been foundational in linguistics. While they do not replace linguistic structure and theory, they serve as model systems and working proofs of concept for gradient, usage-based approaches to language. We offer an optimistic take on the relationship between LLMs and linguistics.

• The paper argues that modern neural language models demonstrate a capability to learn complex syntactic structures, challenging traditional linguistic theories and suggesting a complementary role for LMs in research.
• The success of language models necessitates a reevaluation of overly restrictive hypotheses in linguistic theories and highlights potential parallels with human language processing.
• Integrating language models into linguistics research encourages embracing statistical and functional approaches and promotes interdisciplinary work across computational linguistics, cognitive science, and AI.

How Linguistics Learned to Stop Worrying and Love the LLMs

In recent years, the field of linguistics has been confronted with the rapid advancements in LLMs (LMs), particularly in their ability to produce fluent and grammatical text. This paper, authored by Richard Futrell and Kyle Mahowald, addresses the tensions between traditional linguistic theories and the advent of neural LLMs. It argues that while LMs are not replacements for linguistic theory or structure, they can significantly contribute to and reshape fundamental questions about language structure, processing, and learning.

Central Argument and Methodology

The authors propose a balanced perspective amidst the polarized views on the relevance of LMs to linguistics: some scholars dismiss them as irrelevant due to their architectural differences from human cognition, while others suggest they render linguistic theories obsolete. Futrell and Mahowald assert that both views are extreme and advocate for a middle ground where LMs are seen as complementary to linguistic research.

The paper emphasizes the empirical success of neural LMs in capturing nontrivial linguistic structures, such as subject-verb agreement and recursive syntactic embedding, which were previously thought to be beyond the reach of statistical learning models. This success calls for reevaluating the role of linguistic theories in light of the capacities demonstrated by LMs. The authors utilize historical and theoretical analyses to support their claims and enrich their arguments with findings from various studies on syntactic generalization and structure accumulation in neural networks.

Results and Implications

The analysis reveals several key insights:

1. Nontrivial Grammatical Learning: Modern neural LMs demonstrate the capability to learn complex syntactic structures indicative of linguistic competence, suggesting that their architectural principles may parallel cognitive processes.
2. Challenge to Restrictiveness: The success of LMs necessitates a rethinking of the necessity for overly restrictive hypotheses in linguistic theories.
3. Cognitive and Philosophical Parallels: Just as convolutional neural networks provided insights into human vision, LMs can illuminate aspects of language processing and acquisition, reinforcing the notion of shared foundational principles despite different implementations.

The paper argues that through their ability to model language without pre-programmed rules, LMs underscore the potential of learning systems that operate with fewer constraints than traditionally assumed in linguistic theory.

Future Directions

The authors explore future prospects for linguistics in integrating LMs into research and theory:

• Functional and Statistical Approaches: Encourages embracing functional linguistics and statistical learning paradigms in understanding language evolution and cognitive processes.
• Interdisciplinary Research: Highlights the burgeoning role of interdisciplinary research integrating computational linguistics, cognitive science, and AI to drive forward human understanding of language.
• Bias and Diversity in LLMs: Calls for expanding LM research beyond high-resource languages to include multilingual and under-resourced language contexts.

Conclusion

Futrell and Mahowald’s paper underscores the importance of adopting a nuanced approach to integrating LLMs into linguistic science. By acknowledging the contribution of LMs to our understanding of language, researchers are encouraged to reconceptualize traditional linguistic theories in light of empirical successes achieved through computational approaches. The paper advocates for a collaborative future where LMs and linguistic theories mutually enhance each other, leading to a deeper understanding of human language and cognition.