Unlocking the Past: The Potential of Large Language Models to Revolutionize Transcription of Natural History Collections

Globally, natural history collections (NHCs) house nearly three billion scientific collections offering an unparalleled resource that contributes to both science and society. The first two decades of the twenty-first century have seen a rapid rise in the mobilization of digital biodiversity data. Yet, the transcription of important scientific information from NHCs remains a challenge presenting a critical bottleneck in the digitization process. This experimental study investigates the application of Large Language Models (LLMs) in automating the transcription of herbarium specimen labels. Various LLMs were tested, including OpenAI’s GPT models and Anthropic’s Claude, to evaluate their accuracy and efficiency in transcribing label data. Algorithms have been developed to automate the workflow, leveraging the image processing capabilities of LLMs to bypass traditional Optical Character Recognition (OCR) tools. Prompt engineering strategies were explored to enhance transcription accuracy, focusing on field-specific instructions and formatting. The study demonstrates the potential for LLMs to streamline digitization workflows in natural history collections but underscores several challenges, such as transcription errors, logic errors and inconsistency of results. Promising directions that emerge include improving prompt engineering techniques, exploring the integration of LLMs into workflows and model selection to achieve higher accuracy and reliability.