A Concise Survey of OCR for Low-Resource Languages

Milind Agarwal, Antonios AnastasopoulosNAACL 2024

TL;DR

This paper surveys Optical Character Recognition (OCR) techniques specifically tailored for low-resource languages, which often lack sufficient digitized data for standard NLP tasks. It highlights the challenges and opportunities in digitizing historical documents and proposes a workflow for researchers, linguists, and communities to improve OCR accuracy for these languages.

What Problem Does This Paper Address?

Modern Natural Language Processing (NLP) models thrive on vast amounts of data. However, many languages, particularly Indigenous languages of the Americas, are "low-resource," meaning they have limited digital text and audio available. A significant portion of this linguistic heritage exists in non-machine-readable formats like scanned dictionaries, field notes, and stories. Digitizing these resources using OCR is crucial for preservation and for enabling NLP applications, but standard OCR methods struggle with the unique challenges posed by low-resource languages, such as limited training data, diverse orthographies, and historical document formats. This paper aims to bridge this gap by providing a focused survey of OCR techniques applicable to these specific challenges.

Methods

The paper outlines a four-part digitization pipeline:

  • Preparation: This involves data augmentation to artificially increase the size of limited labeled datasets by applying transformations (e.g., skewing, blurring) to existing images. It also emphasizes pretraining on general unlabeled data and transfer learning from related, higher-resource languages (like Spanish or Portuguese) to initialize models.
  • Training: The focus is on supervised techniques, often starting with a first-pass OCR output from off-the-shelf systems (like Google Vision or Tesseract). Post-OCR processing is then used to correct errors, as it makes fewer assumptions about the OCR system itself.
  • Decoding: Strategies to improve the generation of corrected text are discussed, including copy mechanisms (allowing the model to directly copy words from the input) and lexical decoding (biasing the output towards more frequent or observed word forms).
  • Post-processing: This stage involves refining the OCR output, including text correction to fix errors and coverage mechanisms and diagonal attention to improve the model's focus during generation. The paper also touches upon active learning to strategically select the most informative data for annotation.

Datasets

The paper highlights the Archive of the Indigenous Languages of Latin America (AILLA) as a major repository of undigitized materials. It focuses on 10 specific Indigenous languages of the Americas (listed in Table 1, including S. Bolivian Quechua, Mísqito, Mam, Chuj, etc.) which collectively represent over 5000 pages of undigitized data. While the paper doesn't introduce new datasets for training, it discusses using existing large-scale corpora (like arXiv) for pretraining OCR systems and leveraging related languages for transfer learning.

Key Insights and Ideas

  • Data scarcity is the primary hurdle: For low-resource languages, the lack of labeled data is the most significant challenge, necessitating techniques like data augmentation and transfer learning.
  • Post-OCR processing is critical: Given the limitations of first-pass OCR for these languages, robust post-correction methods are essential to achieve usable text.
  • Leveraging related languages is powerful: Transfer learning from higher-resource languages with similar scripts or linguistic structures can significantly boost OCR performance.
  • Community involvement is key: Successful digitization efforts require collaboration between computational researchers, linguists, and language community members.
  • Layout preservation is an open challenge: Accurately preserving the structural layout of documents remains a difficult problem in OCR, especially for complex historical texts.

Key Results

The paper does not present quantitative results from its own experiments but rather synthesizes findings from existing literature. It highlights that:

  • Data augmentation can increase the robustness and performance of OCR models, especially with small training sets.
  • Copy mechanisms and lexical decoding can reduce character and word error rates by 2-5 times in low-resource settings.
  • Diagonal attention and coverage mechanisms have empirically shown to improve OCR accuracy and sequence-to-sequence performance.
  • Active learning has proven empirically valuable for selecting informative data for annotation in low-resource OCR contexts.

Limitations and Future Work

The authors acknowledge that:

  • A comprehensive guide for all 6000+ low-resource languages is beyond the scope of this paper.
  • Experimental results for the 10 selected Indigenous languages are not presented and would warrant a separate study.
  • Layout preservation is a significant open problem that requires further research, particularly in developing methods for structural post-correction.
  • The effectiveness of transfer learning can vary, and it may require data from a larger set of domains.

Future work could focus on developing more robust layout preservation techniques, exploring advanced pretraining strategies, and creating standardized benchmarks for low-resource OCR evaluation. The paper also suggests that researchers should stay in touch with language communities throughout the process to address any issues and ensure the quality and usability of the digitized data.

Summary updated

BibTeX
@inproceedings{agarwal2024concise,
  author = {Milind Agarwal and Antonios Anastasopoulos},
  title = {{A Concise Survey of OCR for Low-Resource Languages}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
}
Figure