# Large Language Models are not Models of Natural Language: they are Corpus Models

**Authors:** Csaba Veres  
**arXiv:** [2112.07055](https://arxiv.org/abs/2112.07055)  
**Date:** 2021-12-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2112.07055v2/

## Key points

- Natural Language Processing (NLP) is a leading application area in the current Artificial Intelligence boom
- Large deep learning neural networks trained on language modeling tasks have significantly improved performance in downstream language tasks through transfer learning
- Language models trained with data that includes software code can generate functioning computer code from natural language specifications
- Eliminative neural models do not provide any insight into the theoretical foundations of programming languages as they eliminate the need for symbolic abstractions like generative phrase structure grammars
- The term "language model" is misleading and should be replaced with "corpus model"
- The ability of large language models to generate functioning computer code does not necessarily imply a radical restructuring in our understanding of cognition or undermine the symbolic nature of cognitive systems.

## Summary

In the current Artificial Intelligence boom, Natural Language Processing (NLP) has emerged as one of the leading application areas. Large deep learning neural networks trained on language modeling tasks have significantly improved performance in downstream language tasks through transfer learning. Interestingly, when these language models are trained with data that includes software code, they demonstrate remarkable abilities in generating functioning computer code from natural language specifications. However, this raises a conundrum for the claim that eliminative neural models are a radical restructuring in our understanding of cognition as they eliminate the need for symbolic abstractions like generative phrase structure grammars. The syntax of programming languages is determined by phrase structure grammars and thus neural models that produce syntactic code do not provide any insight into the theoretical foundations of programming languages. The demonstration that neural models perform well on tasks involving clearly symbolic systems proves that they cannot be used to argue against the symbolic nature of language and other cognitive systems. As a result, it is argued that the term "language model" is misleading and should be replaced with "corpus model," which better reflects the genesis and contents of these models. In summary, while large language models have shown impressive capabilities in NLP tasks, their ability to generate functioning computer code does not necessarily imply a radical restructuring in our understanding of cognition or undermine the symbolic nature of cognitive systems.