Relevance Theory and LLMs

The Metaphor Between Natural and Artificial Communication: Analogies and Open Issues.

Abstract

According to Relevance Theory, communication is justified by the ability of speakers to mutually activate and recognize intentions and to link them to specific cues or contexts. This is a theoretical model that provides a solution to those phenomena that might potentially undermine communication and comprehension processes: the use of the implicit, so-called figurative language – in this case, the system of metaphors – and, finally, indirect speech acts. From this perspective, Relevance Theory emerges as an effective theoretical model insofar as it traces linguistic-communicative activity back to cognitive processes such as inference and mindreading.

Large Language Models (LLMs) are generative AI linguistic models that understand and generate texts in human language. The activation of LLMs encourages the use of a theoretical model such as Relevance Theory, as can be inferred from the device of prompting.

First and foremost, the activation of LLMs via prompting appears to provide all the elements needed to apply Relevance Theory to LLMs and, consequently, to establish increasingly strong analogies between (natural) human communication and that between AI (which is not cognitively structured!) and human users.

The aim of this paper is to examine whether it is legitimate to speak of artificial mind-reading and whether it is appropriate to restore its natural character by associating it strictly with human users. The analysis of certain ‘case studies’ (the continuum between metaphorical and literal expressions, and the continuum between the implicit and the explicit) may indeed provide some insights into establishing a more precise comparison between artificial communication via LLMs and natural communication. In this regard, it will be useful to examine a specific borderline situation: the role of LLMs in relation to both the translation of dead metaphors from one language to another and the partial handling of new metaphors. The lesser influence exerted by lexicalised elements and the difficulty in finding a broad-spectrum alternative solution highlight certain asymmetries between the two forms of communication, without in any way ruling out the possibility of further investigation.