Gustaaaa/Turkish_Biblio_Ner_Dataset

Dataset for Bibliographic NER from Turkish Book Title Pages

Dataset Summary

This dataset contains annotated images of Turkish book title pages, designed for the task of Named Entity Recognition (NER) to automate bibliographic metadata extraction. It was created for the study titled "The automation of bibliographic cataloging... [Your Paper's Full Title]".

The repository includes two main components:

1. Core Annotated Dataset: A collection of 484 real-world Turkish book title page images that have been manually annotated for key bibliographic entities.
2. Synthetic Augmentation Set: An additional 300 synthetic title page samples generated to investigate the impact of data augmentation on model performance.

Supported Tasks and Leaderboards

The primary intended task for this dataset is Named Entity Recognition (NER). It can be used to train and evaluate models that extract structured metadata (e.g., author, title, publisher) directly from images of book covers and title pages.

Languages

The text in the images and the annotations are in Turkish (tr).

Dataset Structure

Data Instances

Each data instance consists of an image of a book title page and its corresponding annotations, which identify the spans of text for each bibliographic entity.

Data Fields

The annotations identify the following bibliographic entities:

• TITLE: The main title of the book.
• AUTHOR: The name(s) of the author(s).
• PUBLISHER: The name of the publishing house.
• EDITOR: The name(s) of the editor(s), if applicable.
• TRANSLATOR: The name(s) of the translator(s), if applicable.
• (You can add or remove other entities you have annotated here, such as PUBLICATION_YEAR, ISBN, etc.)

Data Splits

The dataset is provided in its entirety to allow for flexibility in experimental design. The original study used this data to train and evaluate several models, including:

• Biblio-BERTurk-NER-Base: Trained on the 484 manually annotated samples.
• Biblio-BERTurk-NER-Extended: Trained on the combined set of 484 real and 300 synthetic samples.

Users are encouraged to create their own training, validation, and test splits as needed for their research.

Dataset Creation

Curation Rationale

This dataset was developed to address the significant challenge of automated bibliographic cataloging for under-resourced languages like Turkish. The goal was to provide a benchmark for evaluating different AI approaches—from fine-tuned domain-specific models to large general-purpose Vision Language Models (VLMs)—on a practical, real-world task.

Source Data

The source data consists of images of title pages from a diverse collection of Turkish books. The synthetic data was generated programmatically to emulate the structure and style of real title pages.

Annotations

The data was manually annotated by the authors. Each entity label was assigned to the corresponding text span found on the title page images. The annotation process was guided by standard library cataloging principles.

Citation Information

If you use this dataset in your research, please cite the original paper:

@article{YourLastName2025_BiblioNER,
  title   = {The automation of bibliographic cataloging, particularly the generation of MARC records from title page images, presents a significant challenge... [Your Paper's Full Title]},
  author  = {[Your Name and Co-authors]},
  journal = {[Journal or Conference Name]},
  year    = {2025},
  url     = {[https://www.reddit.com/r/AskAcademia/comments/snpicp/if_they_publish_a_paper_based_on_tonys_big/](https://www.reddit.com/r/AskAcademia/comments/snpicp/if_they_publish_a_paper_based_on_tonys_big/)}
}
---
license: cc-by-sa-4.0
---