Generative aRgumentative Artificial InteLligence – GRAIL

The GRAIL project aims to study and develop novel, general purpose, generative argumentative AI models, applicable across diverse application domains. Our goal is to create a new generation of AI systems that learn to reason like humans using argumentative reasoning, unlike current generative AI systems. The developed systems will use an explainable argumentation-based reasoning mechanism based on a solid theoretical foundation for building acceptable arguments by applying understandable semantics and leading to explainable and theoretically proven decisions. The proposed reasoning mechanism involves the construction of interactive arguments for or against a belief or action. The reasoner’s decision (i.e., what to believe or do) will rely on the strongest argument, evaluated using a preference relation that compares arguments based on subjective criteria (e.g. beliefs or opinions) or objective criteria (e.g. facts, social or legal norms), along with clear and established acceptability semantics. The reasoner decision depends on the choice of the applied acceptability semantics. This reasoning mechanism ensures that the developed systems will consistently behave as expected. The systems will also be able to explain the trace of their reasoning and the data on which this reasoning is based.
Our project aims to develop generative argumentative AI that combines current advances in language modeling and classical generative AI methods with work in computational argumentation. Thus, we seek to (1) better identify and address the limitations of purely statistical models in reasoning tasks, and (2) bridge argumentation theory with NLP to automatically generate the inputs required for an argumentative reasoning mechanism.