From Situation Model to Word Vector: The Cognitive Limitations of Large Language Models in Processing Pragmatic Ambiguity

Abstract

Understanding language involves more than knowing words and grammar. It also means building a mental picture of the situation
being described. According to psycholinguistic research, readers create and update representations of time, space, causality, intentions, and
characters as they read. This helps them understand not just what is said, but what is meant. Large language models perform well on tasks like
translation and question answering, but they work differently. They learn patterns from large amounts of text, not the kind of rich situational
representations that humans build. Some researchers claim that these models "understand" language, but this view is controversial. This paper
compares humans and LLMs in context processing by examining studies on irony, pronoun resolution, cultural metaphors, and contrast prediction. The findings suggest that LLMs struggle with situational understanding because their mechanisms—word vectors, attention, and nextword prediction—are based on statistical co-occurrence, not on modelling situations. They rely too heavily on word order and distributional
similarity, and not enough on deeper dimensions like time, space, causality, intention, and character. I argue that situation model theory offers
a useful framework for understanding human language processing and for evaluating what LLMs can and cannot do.