Using large language models to extract plant functional traits from unstructured text

Area of Science:

• Ecology
• Computational Biology
• Bioinformatics

Background:

• Functional plant ecology relies on trait data for understanding species distributions and ecosystem functions.
• Existing global trait datasets have limitations, and manual data extraction is time-consuming and expensive.
• Automated extraction of plant traits from textual sources using machine learning is needed to complement field data.

Purpose of the Study:

• To develop and evaluate a natural language processing (NLP) pipeline for extracting plant functional traits from unstructured text.
• To compare the performance of the NLP pipeline against traditional methods like regular expressions.
• To assess the pipeline's effectiveness for both categorical and numerical plant traits.

Main Methods:

• A novel NLP pipeline was designed, incorporating classification models for categorical traits and question-answering models for numerical traits.
• The pipeline was tested on two extensive databases containing over 50,000 species descriptions.
• Various machine learning approaches, including transformer architectures and large language models, were explored.

Main Results:

• The optimized NLP pipeline achieved high performance, with a mean precision of 90.8% and mean recall of 88.6% for categorical traits.
• This represents a significant improvement over regular expression methods, with a 9.83% increase in precision and 42.35% increase in recall.
• The question-answering model for numerical traits demonstrated a normalized mean absolute error of 10.3%.

Conclusions:

• The proposed NLP pipeline effectively automates the extraction of plant functional trait information from diverse textual sources.
• This approach has the potential to greatly accelerate the digitization and utilization of vast amounts of ecological data.
• The findings support the use of advanced NLP techniques to overcome data gaps in functional plant ecology.