Western Foundation of Vertebrate Zoology · Field Notes Transcription Project
This work builds on a Cornell capstone project that explored automated transcription of oological card records from the Cornell University Museum of Vertebrates (CUMV). The current project extends that work to the Western Foundation of Vertebrate Zoology (WFVZ), with a focus on designing and refining a pipeline to transcribe the handwritten and typewritten field notes contained in egg record cards — making this information freely accessible to researchers, particularly for comparative studies on nest construction across species, geography, and time.
The WFVZ has already manually transcribed relevant specimen metadata and scanned card images for their collections. This project focuses specifically on the field notes, which contain rich ecological observations — nest materials, placement, height, substrate, incubation stage — that are not captured in standard museum database fields.
By early April 2026, the CUMV team has processed nearly 3,000 records across three species: American Robin, Gray Catbird, and Northern Cardinal.
Each card set is automatically assigned a triage category based on two signals: the hesitation score (see below) and the content of the transcribed field notes.
| Category | Condition | Meaning |
|---|---|---|
| Red | Hesitation score > 10 | High model uncertainty — review strongly recommended. ~83% chance of transcription error. |
| Amber | Score 3–10 | Moderate uncertainty — spot-check advised, especially for longer notes. |
| Green | Score 0–2 | Low uncertainty — model was confident throughout. <10% error risk. |
| Typed | Card identified as typewritten | Typewritten cards are transcribed with near-perfect accuracy regardless of hesitation score. |
| No Transcript | Field notes blank or header-only | The transcribed field notes field is empty or contains only a label prefix with no real content. |
| Unknown | No logprob file found | No probability data was available for this card set — score could not be computed. |
When GPT-4o transcribes a card, the OpenAI API can return the model's log probabilities for each output token — a measure of how confident the model was in each word or character it chose. A logprob close to 0 means the model was very sure; a more negative value means it was considering alternatives.
For each transcription we isolate the tokens that fall within the
<transcribed_field_notes> span and compute a
hesitation score: the count of content tokens where the difference between
the top-ranked and second-ranked token probability is less than 1.0 — meaning the model
was genuinely uncertain between two plausible readings.
The thresholds (green ≤ 2, amber ≤ 10, red > 10) were calibrated empirically against manually verified transcriptions from the CUMV capstone dataset and refined during the WFVZ processing runs.
Pasch, G. (2024). Automating Transcription of Oological Collection Records Using Large Language Models. Cornell University Master’s Capstone. hdl.handle.net/1813/118249