TDWG2026
Home / Talks / SAL B
talk · Tuesday 22 September · 16:10–⁠16:15 · SAL B

Rediscovering Archives - Using LLM's for the Transcription and Extraction of Ecological Data from Historical Archives

Phoebe Santos · Bots, Bits, and Biodiversity

Recording time 7:31:45–7:36:55Open on Vimeo ↗

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

The short versionOff-the-shelf multimodal LLMs can transcribe varied handwritten archives well, with closed models more robust on structured tables.

Overview

What this was about

Phoebe Santos presented master's research (NHM, ZSL, UCL) testing whether a single LLM transcription approach generalises across different historical record types. Using British Library colonial-era India records and NHM mammal osteology table records, a closed-weight Claude Sonnet model and open-weight Qwen 2.5-VL 32B were prompted to classify page structure (text, mixed, table) and transcribe while preserving structure, with zero-, one- and few-shot and dataset-specific examples. Both did well on text; Claude was significantly better especially on tables and was robust to prompting, while Qwen improved with more and dataset-specific examples.

Why it matters. Evidence that LLM transcription is a viable, affordable route to unlocking historical ecological archives.

Key ideas

In the room

  • Goal: test generalisation of one approach across collections with different handwriting, format and structure.
  • Structure matters for meaning; models often failed to reproduce page formatting.
  • Datasets: British Library 18th-19th century British Colonial Office India records; NHM handwritten mammal osteology tables.
  • Models: Claude Sonnet (closed weights) and Qwen 2.5-VL 32B (open weights), temperature 0.7, no other changes.
  • Prompt asks the model to classify each page as text, mixed or table and apply matching structure instructions.
  • Both good on text; Claude significantly better (small absolute gap) and much better on structured records; Claude consistent across prompting; Qwen improved with more, dataset-specific examples.
  • Claude costs suggested LLM transcription may be cheaper than commercial services; limitations: only English, left-to-right scripts.
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 · 16:05–⁠16:15 · SAL A

Biological occurrence data from historic scientific correspondence: enabling automated approaches for structured data extraction from body text

A Darwin Core-aligned annotated corpus from MfN journals provides a reference for evaluating automated occurrence extraction from historical texts.

Christian Bölling
talk · 22 Sept · 15:00–⁠15:15 · SAL B

Closing Gaps in Specimen Metadata: A Collaborative Annotation and Feedback Platform for Natural History Collections

A national collaborative annotation platform with AI-assisted transcription can close metadata gaps across many Taiwanese collections.

Szu-Hsien Lee
talk · 25 Sept · 09:36–⁠09:48 · SAL A

LLM-Based Pipeline for Extracting Nomenclatural Acts from Taxonomic Literature

A grounded, schema-constrained two-pass LLM pipeline extracts IPNI-ready nomenclatural data well once treatments are found; finding treatment boundaries is the bottleneck.

Ishaipiriyan Karunakularatnam
Programme abstract

Rediscovering Archives - Using LLM’s for the Transcription and Extraction of Ecological Data from Historical Archives

Phoebe Santos

Historical archives held by museums and research institutions contain vast quantities of material that remain underutilised due to limitations in accessibility, readability, and searchability. Recent advances in multimodal large language models (LLMs) present new opportunities for improving access to these collections, particularly for complex historical documents that challenge conventional optical character recognition (OCR) and handwritten text recognition (HTR) approaches. This is especially relevant within natural history archives, where materials range from structured to unstructured, often containing handwritten and typed text alongside historical terminology.

This project investigates the use of multimodal LLMs for the transcription and interpretation of historical archival documents, with a focus on prompt engineering and schema-guided extraction to produce structured, machine-readable outputs. We focus on natural history collection materials such as those housed at Natural History Museum, UK, and Zoological Society London, as well as publicly available datasets such as the India Office botanical records from the British Library. Our study evaluates transcription performance - measured by Levenshtein distance, word error rate, and character error rate - across mixed-format, handwritten documents against ground truth transcriptions. By examining how multimodal LLMs perform on ecologically and historically relevant archival materials, this work contributes to emerging research on AI-assisted archival accessibility and explores how historical collections can be transformed into structured digital resources suitable for biodiversity and historical research.

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