7 Temporal promting for large language models-based on entity linking

Abstract

Large Language Models (LLMs), capable of capturing contextual semantics for accurate disambiguation, have significantly advanced the effectiveness of Entity Linking (EL). However, current methods sometimes suffer from temporal drift, in which the model associates a mention with the most popular or recent entity rather than the accurate entity that was in effect at the time of reference. A methodical investigation of temporal prompting techniques for LLM-based EL is presented in this research. We create four different levels of temporal prompts, from organized templates to implicit time settings, and assess how these affect temporal accuracy. In comparison to baseline prompting, we demonstrate that explicit temporal prompting can reduce drift mistakes by up to 40% using a dataset of temporally-sensitive mentions linked with Wikidata snapshots. The significance of structured temporal signals is underscored by our results, which also point to new avenues for developing temporally resilient EL systems without retraining LLMs.