Neural Generative Models and the Parallel Architecture of Language: A Critical Review and Outlook.

Area of Science:

• Cognitive Science
• Computational Linguistics
• Neuroscience

Background:

• The parallel architecture model posits separate syntactic and semantic processing streams in language comprehension.
• Understanding the interaction between these streams in humans is a key focus in psycholinguistics and neurolinguistics.
• The computational mechanisms underlying this interaction in neural Large Language Models (LLMs) are not well understood.

Purpose of the Study:

• To investigate the extent to which LLMs, specifically GPT-3, can replicate human-like performance on linguistic tasks.
• To evaluate the ability of GPT-3 to process syntactic and semantic information independently.
• To explore the implications of LLM performance for understanding conceptual representation.

Main Methods:

• Revisiting influential linguistic and behavioral experiments.
• Evaluating the performance of the GPT-3 language model on these established tasks.
• Analyzing the model's ability to dissociate semantic from syntactic processing.

Main Results:

• GPT-3 demonstrated the capacity to solve semantic tasks independently of syntactic realization.
• The model's performance on these tasks showed similarities to human behavior.
• The results present a complex and nuanced view of LLM language processing capabilities.

Conclusions:

• LLMs like GPT-3 exhibit human-like semantic processing capabilities, independent of syntax.
• The study highlights the potential of LLMs as tools for investigating language comprehension mechanisms.
• Further research is needed to understand how LLMs acquire structured conceptual representations.