Exploring ChatGPT's Capacity for Modal Subordination Via Discourse Representation Theory

The inherent ambiguity of language presents a challenge to natural language understanding, with modal subordination being one such scenario. The scope of modals, especially those that codify certainty and possibility, gives language the ability to express hypothetical worlds through basic modal logic. This paper will compare the human understanding of modal subordination using Discourse Representation Theory, with the capacity of ChatGPT 3.5 to understand such ambiguity. This is done by asking ChatGPT 3.5 to complete tasks that test its ability to resolve modal subordination, including but not limiting to asking the existence of a discourse referent, and choosing true propositions given a modal scenario. Our results will demonstrate that ChatGPT 3.5 tends to follow the behaviour of modal subordination in clear-cut cases per Discourse Representation Theory, but falls short when pragmatic elements are introduced, or when there exist two modalities.