TDWG2026
Home / Talks / SAL B
talk · Tuesday 22 September · 14:45–⁠15:00 · SAL B

Digital curation across the biodiversity data lifecycle, at the node level (SiB Colombia), aggregator level (GBIF) and how to incorporate user feedback

Esteban Marentes Herrera · Responsible AI, Open Digital Curation, and Round-Tripping for Biodiversity Data

Recording time 6:14:49–6:28:47Open on Vimeo ↗

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

The short versionCuration feedback works at every stage of the data lifecycle, and about a third or more of reported issues do get fixed; AI is used cautiously to summarise feedback.

Overview

What this was about

Esteban Marentes described curation as a circular process with feedback loops at node, aggregator and user levels. At SiB Colombia, every dataset receives human 'acompañamiento' - a semi-automatic review with OpenRefine, the GBIF API and QGIS of metadata and 30-40 Darwin Core elements - before publication. GBIF interprets names and other fields and exposes flags through the Data Validator and annotated downloads. Of about 700 helpdesk messages in three months, 45% of taxonomy issues and about 37% of data-content issues were fixed; an LLM in GitHub Actions summarises taxonomic feedback into JSON (edited by staff about 70% of the time) and a deterministic script checks the Catalogue of Life extended release to close resolved issues.

Why it matters. Quantifies how user feedback flows back to data and shows a careful, human-reviewed use of LLMs in aggregator operations.

Key ideas

In the room

  • Data flow: publisher -> node/data manager feedback -> GBIF feedback -> users providing feedback to GBIF, nodes or publishers; nodes differ in approach.
  • SiB Colombia reviews every dataset before GBIF publication using OpenRefine, the GBIF API and QGIS, recording fixes and required changes in an Excel report; publication only after fixes.
  • GBIF interprets names, elevations, dates and vocabularies and raises flags (fuzzy taxon match, invalid date, invalid coordinates).
  • Publishers can use the GBIF Data Validator; users can see issues in annotated downloads and dataset-level summaries.
  • Roughly 700 helpdesk messages in Q1: mostly general help, then IPT, then taxonomy; 46 about data content.
  • About 45% of taxonomy feedback fixed, largely via the Catalogue of Life extended release; about 37% of data content issues fixed by publishers.
  • An OpenAI LLM in GitHub Actions summarises taxonomic feedback into categorised JSON; staff edit about 70% of outputs; a deterministic script re-checks each issue against the latest CoL extended release monthly.
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 · 14:10–⁠14:25 · SAL A

Reconsidering the original vision for a global biodiversity information facility

GBIF achieved its 1999 vision only partially and through a network of collaborating infrastructures; the community must decide whether that is enough and articulate the next vision.

Joe Miller
talk · 24 Sept · 08:35–⁠08:47 · FORUM

GBIF after 25 years

GBIF sustains itself by leveraging stable national contributions into additional funding and by embedding itself in the biodiversity policy and funding architecture.

Joe Miller
talk · 22 Sept · 11:45–⁠12:00 · SAL C

Modernizing Brazil's taxonomic backbone: how ChecklistBank supports biodiversity and genetic heritage systems

ChecklistBank lets a small national node assemble one consistent taxonomic backbone for legal, portal and permitting systems.

Clara Baringo Fonseca
Programme abstract

Digital curation across the biodiversity data lifecycle, at the node level (SiB Colombia), aggregator level (GBIF) and how to incorporate user feedback

Esteban Marentes Herrera

Digital curation should be a continuous process through the biodiversity data lifecycle. Within the GBIF ecosystem, it starts at the publishing level facilitated by GBIF nodes, continues with quality checks done automatically by GBIF Data Processing and incorporates user feedback provided through GBIF.org or when publishers are contacted directly.

Here we will explain how the data curation is done at SiB Colombia as one example of a GBIF node. Using a publishing support model called “acompañamientos”, where a dedicated helpdesk team reviews some prioritized DwC elements and provides feedback to publishers before publication. With the main goal of facilitating data mobilization by ensuring the data is as complete as possible, meets quality requirements and is free of consistency errors.

The next step is the data processing done automatically by GBIF pipelines before ingestion, addressing some of the most common and relevant elements for users, such as dates, coordinates, controlled vocabularies, etc. These quality flags allow publishers and users to review and improve their data.

The next section covers how the user feedback is currently handled on GBIF, for taxonomic issues that are handled with a human/AI approach, where an AI agent generates JSON tags that are later reviewed by a human team that ensures results are correct and there are no hallucinations.

We also present an overview of user feedback related to occurrence records, covering how much feedback arrives at GBIF, how often publishers respond, and whether they are willing to correct their data.

Finally, we will explore how things can be improved on all fronts to have better data curation and round tripping in the long run for GBIF mediated data.

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