Impact of Knowledge Graph Representations on Question Answering with Language Models

The emergence of Large Language Models (LLMs) brought new approaches to Knowledge Graph Question Answering (KGQA), chasing the vision of querying structured data using natural language. While existing work focuses on improving KGQA approaches, this paper explores the impact of different knowledge graph representations. We consider three dimensions of representation: (i) subsets, (ii) modeling, and (iii) annotations, hypothesizing that different variations impact the F1 scores of KGQA systems. We conduct experiments on a custom knowledge graph featuring integrated data and n-ary relations. Results demonstrate an improvement in the F1 score from 17.6% to 44.5% between the default and best variant, confirming the hypothesis.