A Logical Formalisation of a Hypothesis in Weighted Abduction: towards User-Feedback Dialogues

TL;DR

This paper formalizes user-feedback dialogues in weighted abduction to iteratively refine hypotheses, ensuring the reasoner adapts to user preferences and terminates reliably.

Contribution

It introduces formal protocols for user-feedback in weighted abduction, enabling adaptive hypothesis generation based on user input.

Findings

Protocols guarantee termination under reasonable conditions.

The protocols converge to hypotheses matching user-specified properties.

Applicable across diverse domains like cybersecurity and discourse analysis.

Abstract

Weighted abduction computes hypotheses that explain input observations. A reasoner of weighted abduction first generates possible hypotheses and then selects the hypothesis that is the most plausible. Since a reasoner employs parameters, called weights, that control its plausibility evaluation function, it can output the most plausible hypothesis according to a specific application using application-specific weights. This versatility makes it applicable from plant operation to cybersecurity or discourse analysis. However, the predetermined application-specific weights are not applicable to all cases of the application. Hence, the hypothesis selected by the reasoner does not necessarily seem the most plausible to the user. In order to resolve this problem, this article proposes two types of user-feedback dialogue protocols, in which the user points out, either positively, negatively or neutrally, properties of the hypotheses presented by the reasoner, and the reasoner regenerates hypotheses that satisfy the user's feedback. As it is required for user-feedback dialogue protocols, we then prove: (i) our protocols necessarily terminate under certain reasonable conditions; (ii) they achieve hypotheses that have the same properties in common as fixed target hypotheses do in common if the user determines the positivity, negativity or neutrality of each pointed-out property based on whether the target hypotheses have that property.