Interactive explanations by argumentation

Today’s AI landscape is permeated by plentiful data and dominated by powerful methods with the potential to impact a wide range of human sectors, including healthcare and the practice of law. Yet, this potential is hindered by the opacity of most data-centric AI methods and it is widely acknowledged that AI cannot fully benefit society without addressing its widespread inability to explain its outputs, causing human mistrust and doubts regarding its regulatory and ethical compliance. Extensive research efforts are currently being devoted towards explainable AI, but they are mostly fragmented, tailored to specific settings and focused on engineering static explanations that cannot respond to and benefit from human input.

This PhD project will aim to define a novel scientific paradigm of interactive explanations that can be deployed alongside data-centric AI methods (one or more of recommender systems, or black-box methods or white-box methods) to explain their outputs by providing justifications in their support. This ambitious, cohesive paradigm will be realised using computational argumentation as the underpinning, unifying theoretical foundation: argumentation will be used to provide abstractions for a chosen form of data-centric AI methods from which various explanation types, providing argumentative grounds for outputs, can be drawn, generate conversational explanatory exchanges between humans and machines from interaction patterns instantiated on the argumentative abstractions and explanation type. The novel paradigm will be theoretically defined and informed and tested by experiments and empirical evaluation, and it will lead to a radical re-thinking of explainable AI that can work in synergy with humans within a human-centred but AI-supported society.

Project ID

STAI-CDT-2020-IC-15

Supervisor