SetKE: Knowledge Editing for Knowledge Elements Overlap

TL;DR

This paper identifies a widespread Knowledge Element Overlap (KEO) problem in existing Knowledge Editing (KE) methods, where multiple triplets share common elements and edits can unintentionally overwrite related knowledge. It introduces Knowledge Set Editing (KSE) and a Set Knowledge Editor (SetKE) that uses bipartite matching to align sets of target objects with model predictions, mitigating interference and improving locality. To study KEO rigorously, the authors construct EditSet from Wikidata, comprising 710 relations and over 40k instances across varying numbers of objects, with prompts tailored to evaluate Efficacy, Generalization, and Locality. Experiments across GPT-2 family models and GPT-J show SetKE achieving state-of-the-art performance in KEO scenarios, validating its effectiveness for reliable, scalable knowledge updates in deployed LLMs.

Abstract

Large Language Models (LLMs) excel in tasks such as retrieval and question answering but require updates to incorporate new knowledge and reduce inaccuracies and hallucinations. Traditional updating methods, like fine-tuning and incremental learning, face challenges such as overfitting and high computational costs. Knowledge Editing (KE) provides a promising alternative but often overlooks the Knowledge Element Overlap (KEO) phenomenon, where multiple triplets share common elements, leading to editing conflicts. We identify the prevalence of KEO in existing KE datasets and show its significant impact on current KE methods, causing performance degradation in handling such triplets. To address this, we propose a new formulation, Knowledge Set Editing (KSE), and introduce SetKE, a method that edits sets of triplets simultaneously. Experimental results demonstrate that SetKE outperforms existing methods in KEO scenarios on mainstream LLMs. Additionally, we introduce EditSet, a dataset containing KEO triplets, providing a comprehensive benchmark.

Figures (26)

• Figure 1: A demonstration of normal triplets and KEO triplets, along with a toy example showing how they are localized within Transformer-based LLMs, where normal triplets are mapped to distinct neurons, and KEO triplets are mapped to overlapping neurons.
• Figure 2: Comparison of editing performance on Normal and KEO type for MEND, ROME, and MEMIT.
• Figure 3: Simplified illustration of the SetKE framework.
• Figure 4: Comparing the impact of knowledge overlap number on baselines and SetKE.
• Figure 5: The result of knowledge localization of KEO type knowledge on GPT2.