Libraries are embracing the future, and at the heart of this transformation is a collaborative project between library cataloging experts and data scientists. Recently, I sat down with Synae Yoon, a cataloging librarian with SMU Libraries, and Tue Vu, an AI/Machine Learning Research Scientist with OIT Research Technology Services team, to discuss their innovative approach to streamlining the cataloging of documents and images using Artificial Intelligence into MAchine-Readable Cataloging (MARC) standards.
The journey began when Synae noticed the growing potential of AI in metadata creation. Initially, she experimented with ChatGPT to generate metadata from input images but found the process inefficient—each new image required a fresh prompt. This led to the idea of using a fixed prompt within a chatbot, allowing for a consistent workflow and reducing repetitive manual input. After discussing the concept with her manager and connecting with the Research Technology Services team, the collaboration with Tue was born.
The team explored advanced AI models, including Vision Transformers—a technology that applies transformer models to visual data, enabling the system to analyze images and extract meaningful information, such as captions and physical descriptions. While commercial tools like ChatGPT offered impressive results, they came with limitations: subscription costs, daily usage thresholds, and concerns about data privacy.
To overcome these hurdles, Tue leveraged open-source large language models running on SMU’s high-performance computing infrastructure NVIDIA DGX SuperPOD Advantage. This approach allows catalogers to drastically scale up image processing capabilities securely and efficiently, without the constraints of commercial platforms. Multimodal LLMs like Qwen (from Alibaba) and Gemma (from Google) were used, both featuring robust vision capabilities. All data remained safely housed behind institutional firewalls, ensuring privacy and compliance.
This new AI-powered workflow doesn’t replace catalogers; it empowers them. By generating high-quality first drafts of metadata, the system dramatically reduces the time required to process backlogs and enhances the overall quality of library records. Catalogers still play a vital role in reviewing and refining the metadata, but the repetitive, time-consuming aspects are now handled by AI. This efficiency opens the door to tackling larger volumes of work and improving record accuracy at scale.
The team presented the fruits of this work at a recent Amigos Annual Member Conference, with their session titled: Fixed-Prompt Metadata Workflows: Practical Paths to the AI-Enhanced Library. Collaboration between the Library and OIT has been key, and the team hopes their experience will inspire similar partnerships elsewhere.
