OCR-Based Field Extraction for Archaeological Pottery Metadata: The CENTURIA Dataset
LoRA fine-tuned OCR models convert handwritten pottery records to structured data, reducing error rates to below 1.5%.
Key Findings
Methodology
The study employs LoRA fine-tuning on five OCR models to process 507 pottery records from the Carnuntum site. Fine-tuning on 57 samples significantly improved recognition accuracy.
Key Results
- LoRA fine-tuning reduced transcription error rates from 15-32% to below 1.5%, with field-level accuracy rising above 87%.
- Domain-specific fields were recovered in fewer than 3% of cases without fine-tuning.
- Fine-tuning on 57 samples significantly improved performance on domain-specific fields.
Significance
This study demonstrates how a small expert-validated fine-tuning set can convert handwritten pottery documentation into structured, searchable metadata. This method offers new possibilities for automating archaeological databases, reducing manual transcription workload.
Technical Contribution
By introducing the CENTURIA dataset and LoRA fine-tuning, the study provides new perspectives for automating handwritten document processing, especially in archaeology. This method achieves unprecedented accuracy on domain-specific handwritten records.
Novelty
This is the first use of LoRA fine-tuning on archaeological handwritten records, significantly enhancing OCR model domain adaptability and filling a gap in handling complex handwritten records.
Limitations
- The model still exhibits some error rates in recognizing domain-specific terms and abbreviations.
- The dataset size is limited, which may affect model generalization.
Future Work
Future research directions include expanding the dataset size, exploring applications in other domains, and further improving model performance on complex handwritten records.
AI Executive Summary
Archaeologists often document ceramic finds through technical drawings and handwritten metadata, which are crucial for archaeological research. However, these handwritten records are difficult to analyze computationally, requiring manual transcription. The study introduces the CENTURIA dataset, containing 507 pottery records from the Carnuntum site, providing transcriptions, bounding boxes, and structured field labels. Benchmarking five OCR models revealed a substantial domain gap, with zero-shot transcription errors reaching 15-32% and domain-specific fields recovered in fewer than 3% of cases. However, LoRA fine-tuning on just 57 samples reduced transcription errors to below 1.5% and improved field-level accuracy to above 87%. The study shows that a small expert-validated fine-tuning set suffices to convert handwritten pottery documentation into structured, searchable metadata ready for archaeological databases.
Deep Analysis
Background
Pottery is a primary source for reconstructing past societies. Archaeologists often document ceramic finds through technical drawings and handwritten metadata, which are crucial for dating, provenance attribution, and cross-site comparison. However, these records are difficult to analyze computationally, requiring manual transcription.
Core Problem
Automating the processing of handwritten pottery records is challenging. Existing OCR models perform poorly on complex handwritten records, especially in recognizing domain-specific terms and abbreviations.
Innovation
The study introduces the CENTURIA dataset and LoRA fine-tuning, significantly improving OCR model accuracy on archaeological handwritten records. Fine-tuning on a small sample set greatly enhances performance on domain-specific fields.
Methodology
- �� Benchmarking on the CENTURIA dataset
- �� LoRA fine-tuning on five OCR models
- �� Fine-tuning on 57 samples to optimize model performance on domain-specific fields
Experiments
Experiments were conducted on the CENTURIA dataset, containing 507 pottery records. Five OCR models were benchmarked, and LoRA fine-tuning was performed on 57 samples. Evaluation metrics included transcription error rates and field-level accuracy.
Results
LoRA fine-tuning reduced transcription error rates from 15-32% to below 1.5%, with field-level accuracy rising above 87%. Performance on domain-specific fields significantly improved.
Applications
The study's results can be used to automate archaeological database processing, reducing manual transcription workload and improving data searchability and structuring.
Limitations & Outlook
The model still exhibits some error rates in recognizing domain-specific terms and abbreviations. The dataset size is limited, which may affect model generalization.
Plain Language Accessible to non-experts
Imagine you're in an ancient library filled with handwritten books. You need to convert these books into searchable electronic formats, but the handwritten records are hard to recognize. The study acts like a smart assistant, learning from a small sample set to quickly and accurately convert these handwritten records into structured data, making it easier for you to search and analyze.
ELI14 Explained like you're 14
Imagine you're playing an archaeology game, and you've discovered some ancient handwritten records. These records are important, but you need a tool to help you quickly recognize and organize this information. The study is like a super assistant, learning from a small sample set to quickly and accurately convert these handwritten records into usable data, helping you find clues faster in the game!
Glossary
OCR (Optical Character Recognition)
A technology used to convert text in images into editable text format.
Used to recognize text in handwritten pottery records.
LoRA (Low-Rank Adaptation)
A fine-tuning technique that adapts model parameters through low-rank decomposition.
Used to improve OCR model accuracy in specific domains.
CENTURIA Dataset
An archaeological dataset containing 507 handwritten pottery records.
Used to evaluate and fine-tune OCR models.
Domain-Specific Fields
Terms and abbreviations used in specific domains.
Common in archaeological handwritten records, challenging to recognize.
Fine-Tuning
Adapting a model with a small amount of data.
Used to improve OCR model performance in specific domains.
Open Questions Unanswered questions from this research
- 1 How to validate model generalization on larger datasets?
- 2 Is LoRA fine-tuning equally effective on handwritten records in other domains?
Applications
Immediate Applications
Automating Archaeological Databases
By converting handwritten records into structured data, reduce manual transcription workload and improve data searchability.
Long-term Vision
Cross-Domain Applications
Explore the application of LoRA fine-tuning on handwritten records in other domains, promoting automation in more fields.
Abstract
Pottery is a primary source for reconstructing the chronological and economic dimensions of past societies. Archaeologists often document ceramic finds through technical drawings and handwritten metadata. This metadata is critical for dating, provenance attribution, and cross-site comparison, but remains inaccessible to computational analysis, requiring manual transcription of every record. We investigate whether state-of-the-art document analysis models can address this task, and introduce CENTURIA, a dataset of 507 pottery records from the Roman site of Carnuntum, providing transcriptions, bounding boxes, and structured field-level labels across seven metadata categories. Benchmarking five OCR models reveals a substantial domain gap: zero-shot transcription error reaches 15-32% SpACER-M, far exceeding rates on printed archival documents, with domain-specific fields recovered in fewer than 3% of cases. LoRA fine-tuning on just 57 samples, reflecting a realistic archival annotation budget, closes this gap, reducing transcription error to below 1.5% and recovering overall field-level accuracy above 87%. Our results show that a small expert-validated fine-tuning set suffices to convert handwritten pottery documentation into structured, searchable metadata ready for archaeological databases.