DYNAMICQA: Tracing Internal Knowledge Conflicts in Language Models

TL;DR

This paper introduces DynamicQA, a dataset for studying internal knowledge conflicts in language models, revealing that models struggle with conflicting and dynamic facts, impacting their ability to update knowledge effectively.

Contribution

The paper presents DynamicQA, the first dataset with real-world knowledge conflicts, and evaluates measures to detect intra-memory conflicts in language models.

Findings

Models exhibit more intra-memory conflict with dynamic facts.

Facts with conflicts are harder to update with context.

Retrieval-augmented methods may struggle with conflicting facts.

Abstract

Knowledge-intensive language understanding tasks require Language Models (LMs) to integrate relevant context, mitigating their inherent weaknesses, such as incomplete or outdated knowledge. However, conflicting knowledge can be present in the LM's parameters, termed intra-memory conflict, which can affect a model's propensity to accept contextual knowledge. To study the effect of intra-memory conflict on an LM's ability to accept relevant context, we utilize two knowledge conflict measures and a novel dataset containing inherently conflicting data, DynamicQA. This dataset includes facts with a temporal dynamic nature where facts can change over time and disputable dynamic facts, which can change depending on the viewpoint. DynamicQA is the first to include real-world knowledge conflicts and provide context to study the link between the different types of knowledge conflicts. We also evaluate several measures on their ability to reflect the presence of intra-memory conflict: semantic entropy and a novel coherent persuasion score. With our extensive experiments, we verify that LMs exhibit a greater degree of intra-memory conflict with dynamic facts compared to facts that have a single truth value. Furthermore, we reveal that facts with intra-memory conflict are harder to update with context, suggesting that retrieval-augmented generation will struggle with the most commonly adapted facts.