Leveraging the Power of Complex Logical Query Answering with Natural Language

This study introduces a novel integration of a one-shot complex querying mechanism (ULTRA Query) (Galkin, et al. 2024) with large language models (LLMs) to enhance first-order logic (FOL) querying in knowledge graphs (KGs). We've developed an innovative approach that combines fine-tuning with either the option of using LLMs to function as a semantic parser, translating natural language into FOL expressions, or use the LLM in conjunction with Retrieval-Augmented Generation (RAG) techniques. The expressions generated by either one of the methods described are subsequently converted into Beta Embeddings (Ren, Leskovec 2022), a technique that corrects for the uncertain and dynamic nature of KGs by using probabilistic distributions to predict missing edges. This methodology not only makes it easier to handle the complex and extensive structure of KGs but also works around the challenges posed by their incompleteness, increasing performance by up 25% over the state-of-the-art KG reasoning.