Fully automatic extraction of morphological traits from the Web: utopia or reality?

TL;DR

The paper tackles large-scale extraction of plant morphological traits from unstructured web text by building a three-stage pipeline: harvesting/describing, a description detector trained with a noise-robust loss, and LLM-driven trait extraction via binary trait prompts. Using DistillBERT for description detection and LLM prompts for trait inference, the approach achieves trait values for over half of evaluated species-trait pairs with an average F1 around $0.75$ and coverage near $55\%$–$60\%$ across three real-world datasets. The results demonstrate feasibility for automated construction of structured trait databases from online text, while highlighting limitations due to missing textual descriptions and language diversity. The work suggests that scaling and multilingual expansion, along with improved text coverage, could enable broad, automated trait databases without manual curation.

Abstract

Plant morphological traits, their observable characteristics, are fundamental to understand the role played by each species within their ecosystem. However, compiling trait information for even a moderate number of species is a demanding task that may take experts years to accomplish. At the same time, massive amounts of information about species descriptions is available online in the form of text, although the lack of structure makes this source of data impossible to use at scale. To overcome this, we propose to leverage recent advances in large language models (LLMs) and devise a mechanism for gathering and processing information on plant traits in the form of unstructured textual descriptions, without manual curation. We evaluate our approach by automatically replicating three manually created species-trait matrices. Our method managed to find values for over half of all species-trait pairs, with an F1-score of over 75%. Our results suggest that large-scale creation of structured trait databases from unstructured online text is currently feasible thanks to the information extraction capabilities of LLMs, being limited by the availability of textual descriptions covering all the traits of interest.

Figures (7)

• Figure 1: Overview of the methodology. The panels display the sequence of tasks performed during each of the three main stages: (a) data harvesting, (b) description detection, and (c) trait extraction. Below each task is an example of what its output may look like.
• Figure 2: An example prompt used to query the LLM about the presence of morphological traits given a textual description sentence for a given species (left), along with the LLM response (right). The LLM correctly identifies that there is evidence in the text indicating that the plant type is tree and the phyllotaxis, alternate, while no evidence can be found for the other trait values.
• Figure 3: An example set of sentences and their corresponding "description" score. The text is broken into single sentences by the sentencizer of honnibalspacy2020 and the classifier classifies each sentences. Sentences with a value of 0.50 or higher are stored in the database. The darker the colour green, the higher the prediction value. The prediction value is also shown after each sentence.
• Figure 4: F$_1$-score (orange) and coverage (red) per trait with respect to the three manually curated databases. The coverage is the proportion of species for which at least one value is found. The F$_1$-score is computed only for these species.
• Figure 5: Co-occurrence matrices for every pair of trait values in leaf position and fruit type in the Caribbean dataset, and fruit in the West Africa dataset (left), and the corresponding co-occurrences between the prediction and the annotations (right). We can see that the patterns of co-occurrence is maintained.