Evaluating Complex Entity Knowledge Propagation for Knowledge Editing in LLMs

Abstract

In today's world, where information keeps growing rapidly and changing constantly, language models play a crucial role in making our lives easier across different fields. However, it is tough to keep these models updated with all the new data while making sure they stay accurate and relevant. To tackle this challenge, our study proposes an innovative approach to facilitate the propagation of complex entity knowledge within language models through extensive triplet representation. Using a specially curated dataset (CTR-KE) derived from reliable sources like Wikipedia and Wikidata, the research assesses the efficacy of editing methods in handling intricate relationships between entities across multiple tiers of information. By employing a comprehensive triplet representation strategy, the study aims to enrich contextual understanding while mitigating the risks associated with distorting or forgetting critical information. The study evaluates its proposed methodology using various evaluation metrics and four distinct editing methods across three diverse language models (GPT2-XL, GPT-J, and Llama-2-7b). The results indicate the superiority of mass-editing memory in a transformer (MEMIT) and in-context learning for knowledge editing (IKE) in efficiently executing multiple updates within the triplet representation framework. This research signifies a promising pathway for deeper exploration of data representation for knowledge editing within large language models, and improved understanding of contexts to facilitate continual learning.