ARLtR framework builds knowledge graphs with 19K entities and 8.4K QA pairs

Large language models have dramatically improved information retrieval and question answering, but existing datasets typically force a choice between vector-based retrieval over unstructured text or reasoning over structured knowledge graphs. This binary approach limits the development of hybrid systems that could leverage the strengths of both paradigms. To bridge this gap, researchers Matthijs Jansen op de Haar, Tobias Stähle, and Lorenzo Gatti introduce ARLtR (All Relations Lead to Rome), a unified framework for automated knowledge graph construction and fact-grounded question-answer generation.

The framework jointly builds a knowledge graph, dense embeddings, and question-answer pairs that are explicitly linked to extracted entities, relations, and supporting textual evidence. ARLtR is instantiated as a historical dataset centered on the Roman Empire, comprising over 19,000 entities, 16,000 text chunks, and 8,400 QA pairs. By tightly coupling symbolic graph representations with dense retrieval architectures, ARLtR facilitates evaluation and development of hybrid retrieval systems and semantic steering approaches within a single coherent resource. This enables researchers to benchmark models that can both retrieve unstructured text and reason over explicit relational structures, setting the stage for more robust AI reasoning in knowledge-intensive domains.