LLM Store: A KIF Plugin for Wikidata-Based Knowledge Base Completion via LLMs

Abstract

We present LLM Store, a plugin of a Wikidata-based knowledge integration framework (called KIF) that can be used for knowledge base completion via pre-trained language models in real-time. The LLM Store is one of the participants of the LM-KBC Challenge @ ISWC 2024. This paper describes the implementation of LLM Store, its adaptation to the LM-KBC Challenge, and an analysis of its overall performance — it achieved a macro averaged F1-score of 90.83% in the challenge. One of the key features of LLM Store's approach is a context generation module which uses the external ids and sitelinks in Wikidata to generate textual context for a particular query on-the-fly. The code of LLM Store is available at https://github.com/IBM/kif-llm-store/tree/lm-kbc-challenge.