Large Language Models can extract morphological data from taxonomic descriptions, but their stochastic nature makes automation challenging: a test on Australian Asteraceae

Morphological data are critical for taxonomy, evolutionary biology, ecology, and species identification. However, no widely used central database for morphological data exists as it does for DNA sequences or specimen data. Most of these data are “locked up” in taxonomic literature. Various scripted and Natural Language Processing approaches have been explored to automate the extraction of morphological data from taxonomic descriptions. Here, we explore the feasibility of using Large Language Models (LLMs) and Optical Character Recognition (OCR) to rapidly extract data for 51 morphological characters of Australian native and introduced Asteraceae (daisy family) to populate a taxon × character table. ChatGPT 4o was used to process all 1,121 descriptions, which, following currently accepted taxonomy and after accounting for taxa with descriptions in multiple sources, comprise data for 95 genera and 838 species or infraspecific taxa, totalling 945 taxa. The missing data rate is 51.1%. Visual checking of 109 profiles revealed an error rate of 5.8%, a majority of them misapplication of data to the wrong trait based on confusion between different kinds of bracts and between individual involucral bracts and the involucre as a whole. Error rates were lowest for cypsela and pappus characters, at 2.1%. When repeating 109 inferences with the same LLM, 78.9% of the table cells for which at least one replicate had data showed no substantive difference; the main source of inconsistency was 16.7% of those cells having data in only one replicate. When repeating 109 inferences with an open source LLM run on a local computer, results were considerably less reproducible and showed numerous unit errors, irrelevant information being retrieved, and characters being skipped. Our results suggest that while mining of morphological descriptions with LLMs is possible in principle, instructions for the LLM have to be extremely precise. Even then, in contrast to scripting approaches, LLMs are inherently probabilistic. This makes their responses not fully reproducible and their integration into automated workflows difficult. Future work could explore if results can be improved using approaches such as Retrieval Augmented Generation or fine tuning of models on explanations of morphological terminology. The scripts used in the study and the extracted morphological data for Australian Asteraceae data are made available to support future studies.