💻 computer science

OCR-Based Field Extraction for Archaeological Pottery Metadata: The CENTURIA Dataset

The paper introduces CENTURIA, a dataset of 507 handwritten Roman pottery records, and demonstrates that fine-tuning state-of-the-art OCR models with a minimal set of just 57 annotated samples effectively bridges the domain gap, transforming inaccessible archival metadata into structured, searchable data with high accuracy.

Original authors: Gissu Valentina Naghavi, Dominik Hagmann, Martin Kampel, Irene Ballester

Published 2026-09-01

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Gissu Valentina Naghavi, Dominik Hagmann, Martin Kampel, Irene Ballester

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

History often speaks to us through the things people left behind, but for archaeologists, the most vital clues are frequently hidden in plain sight within their own archives. Pottery shards, broken and scattered across ancient sites, are the primary way researchers understand the daily lives, trade routes, and timelines of past societies. Because clay survives fire and time better than almost anything else, these fragments act as a universal language of the past. However, the story they tell is often locked away in a format that computers cannot read. For decades, archaeologists have recorded these finds by hand, sketching the shape of a broken pot and writing down its details in notebooks or on cards. These handwritten notes contain the essential data: where the piece was found, what kind of pot it was, and how big it is. While the physical drawings might be scanned into a computer, the handwritten text remains just an image, invisible to the search engines and databases that could help historians connect the dots across thousands of years.

The challenge lies in teaching a computer to read these specific, messy notes. Unlike a printed book or a typed document, these archaeological records are filled with unique abbreviations, strange symbols, and handwriting that varies from one archaeologist to another. The notes are often scattered across the page, with no clear order to read them in, and the ink can be faded or crossed out. For a long time, the only way to make this information useful for modern research was to hire people to manually type every single note into a computer, a slow and expensive process that has left vast collections of history inaccessible. The question researchers faced was whether modern artificial intelligence could learn to read these difficult, handwritten archives on its own, turning a static image of a sketch into a searchable list of facts.

A team of researchers from the Technical University of Vienna and the Austrian Archaeological Institute set out to solve this problem by creating a new tool and a new test for the technology. They focused on the massive collection of pottery records from Carnuntum, a major Roman settlement in what is now Austria. This site alone holds roughly 70,000 analogue records, a treasure trove of history that had remained largely unsearchable by computer. The team gathered 507 of these handwritten records to build a dataset they named CENTURIA. Each record in this collection was carefully checked and labeled by human experts, creating a perfect "answer key" that the computer could learn from. This dataset included the original scans, the correct text written out by humans, and a structured map showing exactly which part of the text belonged to which category, such as the location of the find or the type of pottery.

The researchers then tested five different artificial intelligence models to see if they could read these documents without any help. These models are the same kind of technology used to read handwriting in other fields, but when the team let them loose on the pottery records without any training, the results were poor. The computers made mistakes in nearly every single record, misreading letters and numbers so badly that they could not extract the correct information. The error rate was so high that the machines were essentially guessing, failing to recognize the specific abbreviations and symbols that archaeologists use every day. This confirmed that simply applying existing technology to this problem was not enough; the computers needed to be taught the specific language of the archaeologists.

To fix this, the team tried a method called fine-tuning, which is like giving the computer a short, intensive course on the specific subject matter. Instead of showing the machine thousands of examples, which would take years to create, they showed it just 57 examples of the pottery records that had already been corrected by humans. They adjusted the computer's internal settings slightly to help it understand the unique handwriting and abbreviations found in these archives. The change was dramatic. After this brief training, the models stopped guessing and started reading with high accuracy. The error rate dropped from over 30 percent to less than 2 percent. The computers could now correctly identify the location of the find, the type of pot, and the measurements with a success rate of nearly 90 percent.

The study also explored how to make this process reliable enough for real-world use. Since the computers are not perfect, the researchers developed a system where two different models read the same document and compared their answers. If both models agreed on a piece of information, the system accepted it as correct. This simple check allowed the team to automatically trust the results for more than half of the documents, while flagging the few difficult cases for a human to review. This approach means that a small team of experts can now process a massive archive in a fraction of the time it would take to do it manually. The researchers found that with just a tiny amount of expert time to create the initial training examples, they could unlock decades of handwritten records.

The results show that the barrier to digitizing these archives is not a lack of technology, but a lack of specific training data. By proving that a small, carefully curated set of examples is enough to teach a computer how to read these specialized notes, the researchers have opened the door for archives around the world to become searchable. The CENTURIA dataset and the methods developed in this study provide a blueprint for turning piles of old, handwritten sketches into a living, searchable database. This means that historians will soon be able to ask complex questions about the past, such as how pottery styles changed over time or how trade routes shifted, and get answers by searching through millions of records that were previously locked away in ink and paper. The work demonstrates that with the right approach, the stories hidden in the handwriting of the past can finally be heard by the machines of the future.

Technical Summary: CENTURIA – OCR-Based Field Extraction for Archaeological Pottery Metadata

Problem Statement

Archaeological research relies heavily on pottery as a primary source for reconstructing the chronological and economic dimensions of past societies. While digital recording is increasing, the documentation of ceramic finds largely remains an analogue process involving hand-drawn technical cross-sections and handwritten metadata annotations. Even when these documents are retro-digitised, the handwritten annotations—which contain critical excavation context, typological classification, and inventory numbers—remain inaccessible to computational analysis.

Current state-of-the-art Handwritten Text Recognition (HTR) and Key Information Extraction (KIE) models face significant challenges in this domain:

Domain Gap: Standard benchmarks (e.g., IAM, READ) focus on sequential text with canonical reading orders, failing to account for the spatially scattered annotations, faded ink, and dense domain-specific abbreviations found in archaeological records.
Zero-Shot Failure: Pre-trained models struggle with archaeological terminology, leading to high transcription error rates that render raw text insufficient for structured analysis.
Lack of Benchmarks: No existing dataset provides a systematic evaluation for HTR and field extraction on handwritten archaeological pottery documentation, making it difficult for heritage practitioners to assess the viability of automated tools for their archives.

Methodology

The paper proposes a two-stage pipeline to convert scanned pottery documentation into structured, machine-readable metadata: Handwritten Text Recognition (HTR) followed by Field Extraction (KIE).

1. The CENTURIA Dataset

The authors introduce CENTURIA, the first benchmark for this specific task, comprising 507 expert-validated scans of pottery records from the Roman site of Carnuntum (Austria).

Annotation: Each record includes bounding boxes for text and drawings, manual transcriptions, and structured field-level labels across seven semantic categories: excavation (provenance, project, find numbers, etc.), pottery_form, artefact_category, publication_type, surface_treatment, measurement, and others.
Stratification: The dataset captures diverse handwriting styles, scribal conventions, and document conditions (e.g., crossed-out text, overlapping annotations) across six excavation campaigns (1976–2017).
Split: A proportional stratified random split was used: 57 samples for training (fine-tuning) and 450 for testing.

2. Models and Adaptation Strategies

The study evaluates five open-source OCR models spanning different paradigms:

Two-stage detection-recognition: TrOCR (with CRAFT for line segmentation).
Full-page Vision-Language Models (VLMs): olmOCR (1st gen), olmOCR2 (2nd gen, RLVR-trained), Florence2 (generalist), and LightOnOCR (lightweight, 1B parameters).

Three domain adaptation strategies were tested to bridge the gap between general pre-training and archaeological specifics:

Zero-shot Instruction Prompting: Providing domain-specific vocabulary lists and instructions.
Few-shot Prompting: Including $k$ annotated examples ( $k \in \{1, 5, 8\}$ ) in the input context.
LoRA Fine-tuning: Parameter-efficient fine-tuning on the small training set ( $n=57$ ) using Low-Rank Adaptation.

3. Field Extraction

Two extraction methods were compared to map transcribed text to the structured schema:

REGEX-based: A rule-based script using regular expressions and controlled vocabularies (mirroring the ground truth generation process).
LLM-based: Using Qwen3-8B to infer fields from the full transcription sequence via natural language prompting.

Key Results

Transcription Performance

Zero-Shot Limitations: All models exhibited substantial domain gaps, with zero-shot SpACER-M (a metric for scattered text) error rates ranging from 15.58% to 32.35%. Field-level Exact Match Rates (EMR) were critically low (e.g., olmOCR2 achieved only 30.55% EMR), particularly for domain-specific fields like Publication Type and Artefact Category, which often reached near-zero accuracy.
Impact of LoRA: Fine-tuning with just 57 samples dramatically improved performance.
- SpACER-M dropped to <1.5% for both olmOCR2 and LightOnOCR.
- Field-level EMR rose to >89% (92.08% for olmOCR2, 89.40% for LightOnOCR).
- This demonstrates that a minimal, expert-validated annotation budget is sufficient to overcome the domain gap, outperforming zero-shot prompting and few-shot strategies (the latter of which caused LightOnOCR to enter repetitive output loops).

Field Extraction Performance

Transcription as the Bottleneck: Extraction accuracy was heavily dependent on transcription quality. Once LoRA fine-tuning improved the text, extraction accuracy followed, with no significant difference between REGEX and LLM extractors on fine-tuned inputs (both achieving ~90% overall accuracy).
Error Modes: The primary failure mode remained character-level OCR errors (e.g., digit substitution in dates or measurements) rather than extraction logic failures. Short domain-specific tokens (e.g., surface treatment abbreviations) were prone to omission under REGEX, while LLMs tended to hallucinate incorrect values for these tokens.

Practical Deployment Heuristics

The authors propose a cross-pipeline confidence estimation strategy to automate archival processing:

By running four pipeline combinations (two models $\times$ two extractors), fields where all four agree are correct 98.7% of the time.
Using a "Smart Merge" heuristic (majority vote and rule-based resolution for conflicts), 56.4% of documents can be auto-accepted with 95.5% field precision.
This allows heritage practitioners to concentrate manual review efforts only on the remaining ~44% of documents where pipeline disagreements persist.

Significance and Claims

The paper claims that CENTURIA provides the first systematic benchmark for HTR and field extraction on handwritten archaeological pottery documentation. Its primary contribution is demonstrating that automated conversion of decades of handwritten documentation into searchable metadata is feasible with a realistic annotation budget.

Key takeaways emphasized by the authors:

Small Data Sufficiency: High-accuracy domain adaptation does not require massive datasets; fine-tuning on ~57 expert-validated samples is sufficient to reduce transcription errors from >15% to <1.5%.
Pipeline Robustness: While zero-shot models fail, a combination of LoRA fine-tuning and cross-pipeline verification creates a viable workflow for large-scale archival digitisation.
Accessibility: By releasing the dataset, checkpoints, and code under CC-BY-4.0, the authors aim to enable heritage practitioners to make informed decisions about deploying these tools on their own archives, moving from inaccessible visual records to structured, computable data.

The authors remain modest regarding generalizability, noting that while the challenges (scattered annotations, abbreviations) are common globally, cross-archive generalization depends on specific recording conventions, scripts, and languages, and requires further evaluation.