Symbols and grounding in large language models

Abstract

Large language models (LLMs) are one of the most impressive achievements of artificial intelligence in recent years. However, their relevance to the study of language more broadly remains unclear. This article considers the potential of LLMs to serve as models of language understanding in humans. While debate on this question typically centres around models’ performance on challenging language understanding tasks, this article argues that the answer depends on models’ underlying competence, and thus that the focus of the debate should be on empirical work which seeks to characterize the representations and processing algorithms that underlie model behaviour. From this perspective, the article offers counterarguments to two commonly cited reasons why LLMs cannot serve as plausible models of language in humans: their lack of symbolic structure and their lack of grounding. For each, a case is made that recent empirical trends undermine the common assumptions about LLMs, and thus that it is premature to draw conclusions about LLMs’ ability (or lack thereof) to offer insights on human language representation and understanding. This article is part of a discussion meeting issue ‘Cognitive artificial intelligence’.

Neural Networks as Cognitive Models of the Processing of Syntactic Constraints

“…In particular, to match the hypothesized mechanism underlying human behavior for the distributivity experiments (Humphreys & Bock, 2005 ), a model would need to distinguish between, for example, an NP that is more likely to be conceptualized as a single, collective entity and an NP that is more likely to be conceptualized as multiple entities distributed in space. This kind of mapping, from linguistic material to entities in an external world, may lie beyond the abilities of models trained solely on linguistic material at this scale (though see Pavlick, 2023 for evidence that these capacities may emerge when models are trained on orders of magnitude more training data). We speculate that a multi-modal model with a visual training objective may be better able to capture such effects (for a example of a multi-modal model in distributional semantics, see Bruni et al, 2014 ).…”

How could we make a social robot? A virtual bargaining approach

“…Various reasons for scepticism have been advanced about the significance of GPT-3's impressive performance. Pavlick [ 19 ] argues that certain well-known sceptical concerns are not yet decisive. In particular, she makes the case that objections that large language models like GPT-3 cannot really understand language because they lack internal symbolic structured representations of language, or that their symbols are not ‘grounded’ in perceptuo-motor interaction with the world, are currently unproven.…”

Argument and explanation

“…This makes more detailed analysis of current capabilities rather futile. There are, however, interesting questions about the relationship between AI and cognitive science that are posed by these models, and we refer the reader to two papers in this special issue that pursue these further [39,40].…”