Grant: Discovery Launch Supplement

The overarching research goal of this proposal is to characterize how modern artificial intelligence systems represent and manipulate natural language meaning. These systems rely on large language models (LLM) trained on vast amounts of textual data to produce representations of natural language expressions. The backbone of our project will be to tease apart how these representations capture different facets of meaning, focusing on parts of meaning that are not inherently objective. In particular, we plan to evaluate the ability of such systems to manipulate the argumentative dimension of language, i.e. to use linguistics means for persuasion. A number of approaches in theoretical linguistics propose that natural language meaning is inherently argumentative: the linguistic code offers expressions and constructions that shape how of an utterance can be used to argue for certain conclusions, in a way that is largely independent from the objective reality described by the sentence. This argumentative dimension is notably manifested in the meaning of lexical items and in the semantics of certain discourse markers and connectives. All of these elements assume that the speaker of an utterance speaks with a specific purpose in the light of which the utterance is interpreted. The argumentative potential of a linguistic expression is inherently distributional: if two expressions can be substituted for each other in a given context without altering the overall cohesion and coherence of a discourse, we assume they share some of their argumentative potential. These distributional aspects dovetail with the information used by LLM to form the representations they offer ; we thus expect LLM to capture argumentative content efficiently. Concretely, we will investigate how LLM representations encode argumentative content at the level of lexical items and argumentative operators, and compare our findings across three unrelated languages: French, English and Written Chinese. We expect to observe cross-linguistic similarities at the level of argumentative operators, which would provide further evidence to the hypothesis that natural language encodes an argumentative component. The research has implications in theoretical and computational linguistics: it will help sharpen and bring an empirical basis to existing theories of argumentation within language, and will also help to characterize the semantic capabilities of LLM. In particular, it will test the hypothesis that these models are trained to be sophists, i.e. systems whose goal is to produce coherent and cohesive discourse on any thesis, with no regards for truth. The research has thus implications for the way such systems are used for the general public, and discussed in the public debate, in particular in the media.