Header

Suche
Digital Society Initiative

AI-R: Using AI to Navigate the Space of Reasons

It is a characteristic of human beings that we play the game of giving and asking for reasons (Brandom 1994; 2021). We continuously ask for, give, and reflect on reasons. In doing so, like intellectual spiders, individuals, communities, and companies weave individual reasons, that is, propositions, into webs that constitute one’s explicit framework from which to accommodate and judge new claims with which one is confronted. The threads that give coherence to this web are logical relationships of entailment. What gives dynamics to the web are contradictions. 

For example, let us assume that I believe that it is not possible for dogs to get along with cats. This proposition entails that my own cat Carl cannot get along with my neighbor’s dog Rudy. Then, however, I see the two peacefully sitting on a riverbank, watching the river flow. This seems to contradict my specific belief about the possible relationships between cats and dogs, forcing me to modify either this belief, and hence the more general belief about cats and dogs, or to find an interpretation of the behavior of the two animals that does not amount to the two getting along well with each other, which would then potentially require adaptations elsewhere in my web of belief, in particular about what follows from two beings sitting peacefully next to each other. This metaphor of the web of belief and its dynamics was brought into the philosophical debate by W.V.O. Quine, and it has been developed further by inferentialists, pioneering among them Robert Brandom.

Being aware of what one's beliefs presuppose and entail is a core constituent of rationality: It allows an individual, a political party, or a company to be conscientious of their position's presuppositions, of its common ground with political or ideological opponents, and it facilitates the identification of inconsistencies within one's own position. While the basic ideal of better understanding one's web of belief has accompanied western philosophical thought in one form or another since Plato's dialogues (the Socratic method being essentially a means to identify inconsistencies within a given web of belief), we believe that our time holds particular promise for progress towards this ideal, given progress both in philosophical-conceptual developments of the past decades as well as in NLP over the past 8 years:

  1. On the philosophical side, there is the painstaking detail and acumen with which Robert Brandom and other inferentialist philosophers have described the inferential web of belief that societies and individuals weave . 
  2. In NLP, with the advent of generative large language models based on modern deep learning (DL) such as the Transformer models, the time is right to put these philosophical concepts to work and hence to test. Throughout, we follow a radically open approach, encompassing open science, a levelling of the playing field in terms of computational resources, and in the way we interact with the general public that funds our research.

The project pursues three main goals:

  1. Combining insights from neuro-symbolic NLP and inferentialism, we will build a new method that sets a new state of the art in identifying the logical relationship both claimed and actually existing between propositions. In NLP, this encompasses aspects of two fields, namely Argument Mining and Natural Language Inference (NLI).
  2. Using this method and grounded in actual data, we will build a representation of real-world webs of belief of both individuals and legal entities such as political parties or companies.
  3. We will use insights from creativity research, that is, aesthetics, to insightfully surprise by identifying regions in these webs that can ground further development, perhaps because they contain logical inconsistencies, or because they constitute isolated nodes not grounded in further propositions.

 



Projektlaufzeit: 4 Jahre, Projektstart im Juni 2024

Projektleitung: