Bridging Open Science and large language models: Enhancing research accuracy through Knowledge graphs

This poster was presented at the Open Science Conference 2025, hosted by Leibniz Strategy Forum Open Science and organised under the lead of the ZBW – Leibniz Information Centre for Economics.

Open Science (OS) emphasises reproducibility, factual accuracy and originality to promote responsible conduct of research, share reliable data, minimise resource waste, and foster innovation.

In contrast, large language models (LLMs) process vast amounts of (non-) scientific data by probabilistic modeling, prioritising quantity over reliability of data.

As LLMs are increasingly used in research, a critical question arises: How can the two different logics (1)—efficiency through openness and efficiency through volume—coexist and be utilised responsibly in research? We expand this question and argue for knowledge-augmented systems to enhance LLMs’ accuracy and reliability, proposing a combination with an OS knowledge graph (KG).

The poster visualises a triangular relationship among OS, KG, and LLM, highlighting two workflows.

(A) OS—KG—LLM: OS resources serve as the knowledge base structured within a KG ontology (2), connected upstream of an LLM, providing a constantly updated reference that guides the LLM’s data selection through specific nodes (entities) and branches (relationships), mitigating hallucinations and delivering contextually relevant content aligned with the prompt.

(B) OS—LLM—KG: OS resources are automatically prepared with the help of an LLM, which identifies nodes and relationships, extracts relevant data from the OS pool, and transforms publications into KG-compliant data. Thus, LLMs can assist in creating an ontology represented in a KG based on the principles, data, and results of OS.

References:

(1) https://doi.org/10.5281/zenodo.11562117 

(2) https://doi.org/10.48550/arXiv.2406.08223