At UC3, several of our latest initiatives involve integrating AI tools, with a particular focus on improving metadata and assisting researchers with creating best practice DMPs.
A clear philosophy guides UC3’s approach to the use of generative AI: addressing researchers’ and the broader research community’s needs, keeping humans as the authority, complementing human work for scale and efficiency, and prioritizing open-source solutions where possible.
Improving ROR Metadata
One key application of AI we are exploring is enhancing the quality and scale of our metadata curation activities, including those for the Research Organization Registry (ROR). ROR, a widely adopted persistent identifier service for research organizations, operates on a model where anyone can submit a request to add or update its records. This community-focused approach to curation has allowed ROR to grow rapidly by gathering diverse and valuable feedback from a global userbase. However, as one might expect with crowd-sourced data, it also has inherent complexities that require special attention to maintain consistency and quality.
AI helps by taking these diverse user inputs and automatically transforming them into clean, structured, authoritative outputs in the ROR dataset. For adding records to the registry, this automation seamlessly handles data standardization, formatting, and enrichment tasks that would otherwise require specialized logic and manual intervention to achieve. For updates to the registry, AI can transform natural language descriptions of desired changes into structured modifications, described using ROR’s data model. These interventions have dramatically accelerated ROR’s request processing ability, enabling the service to now efficiently handle its growing request volume and process over 1,000 user-submitted requests per month.
Despite these advances, achieving 100% accuracy or completeness with these methods is neither possible nor desirable. Instead, we choose to pursue hybrid approaches that balance the efficiency and scalability of GenAI with the measured judgment and domain expertise that only human curators can provide. In doing so, we can embrace both innovation and authoritative oversight, allowing ROR to further grow in its position as a reliable, community-driven infrastructure, in service to the complex needs of the global research ecosystem.
DMP Chef: Exploring AI-powered DMP Generation
Another example of our AI exploration is “DMP Chef,” a large language model (LLM) based DMP generator. We are in the initial stages of this work, partnering with the California Medical Innovations Institute (CalMI2) to develop a new tool that allows researchers to provide simple descriptions of their work, from which the DMP Chef can generate a draft DMP. We are currently developing this tool to work with NIH DMPs and plan to follow up on this work by working with NSF and other templates.
The current process involves asking researchers for a short description of their study and the types of data they plan to collect, then using a detailed prompt to have the LLM draft an initial DMP using NIH’s template for review. To test the initial quality, we used the NIH exemplar DMP, extracting the study design and data types from Element 1, and then feeding that information into the tool. We compared the generated output with the actual DMP section by section. Our next step is to recruit data librarians to review these generated DMPs for quality and comprehensiveness.
We’re seeing some moderate initial success with off-the-shelf LLM models, including open-source models, and plan to continue working on refining the quality by exploring options such as asking additional questions to the researcher, generating sections to separate, and feeding the LLM additional policy documents. Our goal is to help create an initial draft of a high-quality plan that researchers can then refine to their needs, suggesting best practice repositories and standards based on their specific data.
Matching Related Works: Connecting Plans to Outputs
We’re also developing new tools to automatically connect DMPs to the research outputs they describe, such as datasets, articles, and software. These new connections improve the discoverability of research data and make it easier for researchers, funders, and administrators to see the complete picture of a project’s outputs. Our approach combines structured metadata from maDMPs with information from sources like DataCite, Crossref, OpenAlex, and the Make Data Count Citation Corpus. We utilize machine learning, incorporating embeddings generated by large language models and vector similarity search, to compare the text from the title and abstract of a DMP with those descriptive fields within the datasets, rather than relying solely on metadata for authors and funders. A human reviewer then confirms the matches to ensure accuracy and reduce the manual reporting burden on researchers. You can read more about this feature at the DMP Tool Blog.
UC3’s AI initiatives are focused on making research data easier to find, connect, and trust. By pairing AI-driven efficiencies with human expertise, we can accelerate workflows while maintaining the accuracy, transparency, and trust essential to research.
