M-Rational: Reconceiving and Improving Multi-Perspectivality and Rationality in Argumentative AI
With the introduction of the transformer architecture Vaswani et al. (2017) and its recent evolution into Large Language Models (LLMs), a number of approaches developed in Natural Language Processing (NLP) reached a level of performance that makes them ready (or almost ready) for deployment in real-world settings. Many of these applications are evolving from a text classification perspective (e.g. sentiment analysis, information extraction) to complex reasoning tasks including in complex domains of discourse. Some of these domains of discourse, such as political, ethical and legal reasoning, requires an interpretation model which not only models processes of inference, both formal and material, but also requires the acknowledgment and integration of multiple assumed perspectives and value systems.
The dominant paradigm for the construction of these systems involves the notion of a human-grounded feedback, the use of annotations, instructions or post-hoc validation for the training or evaluation of these models. This paradigm, when applied to complex and multi-perspective domains, bring the inherent risk of introducing severe subjective biases into these models. As generative language models march towards widespread application, developing new methodologies for integrating multi-perspective reasoning capabilities into these models becomes a matter of strategic societal importance.
This project focuses on the development of new conceptual resources and methods to support the development of Natural Language Inference (NLI) models which deliver multi-perspective reasoning. To achieve this ambition, it is conceptualized across four main pillars:
- a novel conception of multi-perspective reasoning based on inferentialism (MPR),
- a novel conceptual understanding of perspectival rationality inspired by the philosophy of science,
- the development of novel mechanisms for controlling LLMs in order to deliver MPR and
- new evaluation methodologies for MPR models.
In NLP, there has been a long-standing orthodoxy of unifying disagreeing perspectives – be it in the creation of a training dataset or in the evaluation of the output of LLMs – into one single ground truth, usually by majority vote. More recently, however, researchers have started to explore ways of appreciating and exploiting the plurality of perspectives (what we call perspectivality) that can become visible in the context of challenging tasks such as the computational assessment of argument quality.
However, the basic issue with this positive conception of perspectivality is that, to this day, it suffers from both a theoretical and a practical deficiency. On the theoretical side, it is not clear (i) how exactly to understand the notion of a perspective, including basic aspects such as the question of what such perspectives are perspectives of, how to give an exact account of such perspectives, and what makes such perspectives rational positions in the space of reasons (rather than simple aberrations or noise).
M-Rational will develop conceptual answer to these basic questions. Our project aims to chart new terrain, neither forcing everything into the Procrustean bed of a single unified ground truth, nor collapsing into a debilitating «anything goes» relativism. On the practical side, M-Rational will contribute to evolving the state of the art in NLI and argumentation reasoning by shaping the sub-theme of multi-perspective argumentation reasoning, i.e. models which can interpret and critically augment arguments, mapping representative stances, eliciting underlying premises (assumed facts, definitions) and maximize logical validity. Combining with the philosophical part, the project will also modify the very notion of what it means to make progress at argument reasoning in multi-perspectival settings by reconceiving the goal of the task as such.
Projektleitung
- Dr. Reto Gubelmann (PI – DSI, Universität Zürich)
- Prof. Dr. Christina Niklaus (University of St.Gallen)
- Dr. André Freitas (IDIAP)
Finanzierung
Schweizerischer Nationalfonds (SNF)