TDWG2026
Home / Talks / SAL A
talk · Friday 25 September · 10:54–⁠11:06 · SAL A

Curation through citation: using AI and a knowledge graph to curate DNA barcodes

Roderic Page · Large Language Models for Biodiversity Data Discovery, Integration, and Curation. AI applications, integration, and validation

Recording time 3:24:59–3:35:04Open on Vimeo ↗

Summaries and key moments are machine-produced. Check the recording for the speaker’s exact words.

The short versionLinking barcodes, specimens and literature in a knowledge graph could fix many BOLD records, if we can reliably extract the connecting codes from papers.

Overview

What this was about

Roderic Page asked what concrete value knowledge graphs deliver, and proposed curation of DNA barcode data as an answer. Using his BOLD View site, he showed BOLD records with bad or missing coordinates, vague or informal names, and names hidden behind GenBank links. Some can be fixed by following links from the barcode to GenBank, the publishing paper's DOI and a holotype, which is more than internal-consistency data cleaning. The barriers are extracting specimen and sequence codes from literature (which Plazi does not do well) and linking barcodes to papers. He pointed to emerging large triple stores (Wikidata, OpenStreetMap, GBIF, BHL, Bionomia; 86 billion triples) as signs that such a graph is becoming real.

Why it matters. Barcode data train widely used identification models while many records lack species names or correct coordinates. Citation-based curation would improve data that downstream AI relies on.

Key ideas

In the room

  • Knowledge graphs need to show 'concrete information you do not already possess' (the Silicon Valley quote).
  • CurateGPT suggested combining knowledge graphs and chatbots for curation.
  • BOLD holds ~25 million barcodes; a well-known 5-million-barcode training set is mostly identified only to order or family.
  • BOLD View is his site for exploring barcode data and noting odd records.
  • Examples: a Tasmanian-endemic crustacean plotted at Australia's centroid; a record labelled only 'Caudata' that GenBank calls a snake from China; an informal name resolved via GenBank -> DOI -> paper -> holotype.
  • Data cleaning flags inconsistencies; fixing needs external evidence such as the literature.
  • Plazi extracts coordinates well but not specimen and sequence codes; a simple ChatGPT prompt gets most of them.
Jump in

Notable moments

In their words

Transcript

Automatically generated captions can contain mistakes, especially in names and technical terms. Times are relative to the room recording.

Read the transcript ↓
Loading transcript…
Keep exploring

Related talks.

See the map ↗
talk · 22 Sept · 14:40–⁠14:55 · SAL A

BHL and the Planetary Knowledge Base: Structuring Biodiversity Literature for the Next Generation

NHM is re-OCRing all of BHL with modern models and extracting entities into a knowledge graph, with confidence gating to protect BHL data quality.

Qianqian Hiris Gu
talk · 22 Sept · 14:00–⁠14:15 · SAL B

Open Digital Curation and Round-Tripping Biodiversity Data Enhancements Across Collection Infrastructures: Six Years and 50 Million Links in Bionomia

Round-tripping is much harder than expected, and data managers must fully populate identifiedBy so taxonomists' expertise is credited.

David Shorthouse
talk · 24 Sept · 14:38–⁠14:50 · FORUM

Building digital-native species: Transitioning from PDF liberation to semantic-by-design publishing

Keeping data linked and structured from specimen to publication yields FAIR, trusted data on the day of publication.

Laurence
Programme abstract

Curation through citation: using AI and a knowledge graph to curate DNA barcodes

Roderic Page

Recent work on visualising DNA barcode data in the BOLD database withBOLD View has revealed numerous records that are problematic or incomplete. Common problems are a lack of formal identification ("dark taxa" in the original sense), a lack of geographic coordinates, and missing information on type status.

Often the quality of a barcode record can be improved by consulting the paper that published the DNA barcode sequence. Papers frequently contain details on sampling localities, and an informally named taxon may subsequently have been described. For example, the barcode LEASZ362-22, published in the dataset "Lepidoptera of Cottian Alps" https://doi.org/10.5883/DS-LEPICOTT, is identified only to genus level as "Denisia". The DS-LEPICOTT dataset is cited by Huemer et al. (2023) https://doi.org/10.3390/d15020214, which in turn is cited by Huemer (2024) https://doi.org/10.3897/alpento.8.127794, which describes the new species, Denisia cryptica, for which the barcode LEASZ362-22 is from the holotype. Following this chain of connections resolves the identity of a dark taxon.

For this kind of curation to be feasible at scale we need to be able to discover these connections automatically. One way to do this is to create databases of the intermediate connections, especially citations between papers, and citations of data. Some of these links (such as papers citing papers) are available in existing databases, others have to be assembled by text mining or record matching.

This talk describes constructing this knowledge graph, and the role of AI tools both in constructing the graph and in enabling users to query the graph using natural language. Then talk will argue that a richly connected knowledge graph has the potential to improve the quality of biodiversity databases.

Submitted to the conference programme before the talk, so the talk itself may differ.