“To Have the ‘Million’ Readers Yet”: Building a Digitally Enhanced Edition of the Bilingual Irish-English Newspaper an Gaodhal (1881-1898)

Oksana Dereza, Deirdre Ní Chonghaile, Nicholas WolfCOLING 2024

TL;DR

This paper introduces novel OCR and post-correction models for the historically under-resourced Irish language, specifically for the bilingual newspaper An Gaodhal. By developing these tools and releasing the dataset, the project aims to unlock access to a rich historical resource and pave the way for similar efforts for other endangered languages.

What Problem Does This Paper Address?

Digitizing and making historical texts accessible is crucial for research, but it's particularly challenging for languages with limited digital resources. The Irish language, or Gaeilge, falls into this category. The authors tackle the problem of accurately extracting text from An Gaodhal, a bilingual Irish-English newspaper from the late 19th century. This newspaper presents unique challenges due to its use of a specific Irish typeface (Cló Gaelach), pre-standardized spelling, and code-mixing between Irish and English. Existing OCR tools are often not equipped to handle these complexities, leading to high error rates and hindering research.

Methods

The core of the project involves developing and fine-tuning OCR and post-correction models.

  • OCR Model Development: The team trained two new OCR models:
  • One for Irish text in the Cló Gaelach typeface.
  • Another for bilingual Irish-English content, accommodating both Cló Gaelach and Roman typefaces.
  • OCR Post-Correction: A key innovation is a BART-based OCR post-correction model. This model was fine-tuned on the An Gaodhal data to specifically address errors common in historical Irish-English texts.
  • Named Entity Recognition (NER): The project also created a historical Irish training set for NER, aiming to automatically identify entities like people, places, and organizations within the text.
  • Data Augmentation and Balancing: To improve model performance, the team experimented with techniques like masking, dataset balancing (adjusting the ratio of correct to incorrect examples), and data augmentation (repeating lines).
  • Manual Correction: A significant manual correction effort was undertaken to create high-quality training data and validate the automated processes.

Datasets

The primary dataset used is the digitized collection of the bilingual Irish-English newspaper An Gaodhal (1881-1898). This newspaper is described as a unique resource due to its historical context, bilingual nature, and specific Irish typeface. The authors also mention using a Cló Gaelach training dataset for Tesseract OCR software as a related existing resource.

Key Insights and Ideas

  • The Value of Specialized Models: Generic OCR models struggle with the unique challenges of historical Irish texts (typeface, spelling, code-mixing). Developing specialized models is essential for accurate digitization.
  • BART's Efficacy for Post-Correction: The BART model demonstrated significant promise for OCR post-correction, outperforming other transformer models tested.
  • The Importance of Data Quality: Manual correction and careful dataset balancing were critical for achieving good results, highlighting that even with advanced models, high-quality training data is paramount.
  • Historical Context Matters: Understanding the historical context of printing, language standardization, and usage conventions (like abbreviations) is crucial for both OCR and manual correction.
  • Potential for Endangered Languages: The methodologies and tools developed have the potential to be adapted for other under-resourced and endangered languages, democratizing access to their historical records.

Key Results

  • The project successfully trained two new OCR models for Irish and bilingual Irish-English content.
  • A BART-based OCR post-correction model was developed, achieving improved accuracy (lower CER/WER) compared to raw OCR output. The best configuration achieved a test CER of 3.27% and WER of 10.17%.
  • A historical Irish training set for Named Entity Recognition (NER) was created.
  • The team released the machine-readable full text of An Gaodhal in ALTO-format XML files.

Limitations and Future Work

  • Manual Correction Bottleneck: While automated methods were employed, manual correction was still a significant part of the process, especially for complex cases or where automated models failed.
  • Limited NER for Irish: The NER work for Irish is still in its early stages, with a dataset of 11,000 words labeled for training.
  • Future Work: The team plans to continue working on NER for Irish, explore applying these tools to other historical Irish newspapers like An Stoc, and make all project outputs publicly accessible for further research. They also hope their methodologies can inform efforts for other under-resourced languages.

Summary updated

BibTeX
@inproceedings{dereza2024have,
  author = {Oksana Dereza and Deirdre Ní Chonghaile and Nicholas Wolf},
  title = {{“To Have the ‘Million’ Readers Yet”: Building a Digitally Enhanced Edition of the Bilingual Irish-English Newspaper an Gaodhal (1881-1898)}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2024},
}
Figure