Rich Knowledge Sources Bring Complex Knowledge Conflicts: Recalibrating Models to Reflect Conflicting Evidence

TL;DR

This paper investigates how question answering models handle conflicting information from large knowledge sources, revealing reliance on non-parametric data and minimal confidence adjustment in contradictions, and proposes a calibration method to improve response consistency.

Contribution

The study highlights the impact of knowledge conflicts on model reliance and introduces a calibration approach to better manage conflicting evidence in question answering.

Findings

Models rely more on retrieved passages than parametric knowledge.

Contradictions among sources have little effect on model confidence.

Calibration discourages models from committing to a single answer amidst conflicts.

Abstract

Question answering models can use rich knowledge sources -- up to one hundred retrieved passages and parametric knowledge in the large-scale language model (LM). Prior work assumes information in such knowledge sources is consistent with each other, paying little attention to how models blend information stored in their LM parameters with that from retrieved evidence documents. In this paper, we simulate knowledge conflicts (i.e., where parametric knowledge suggests one answer and different passages suggest different answers) and examine model behaviors. We find retrieval performance heavily impacts which sources models rely on, and current models mostly rely on non-parametric knowledge in their best-performing settings. We discover a troubling trend that contradictions among knowledge sources affect model confidence only marginally. To address this issue, we present a new calibration study, where models are discouraged from presenting any single answer when presented with multiple conflicting answer candidates in retrieved evidences.