---
title: "Arabic Handwritten Text Line Dataset — Bouchal 2023"
description: "The paper addresses a gap in resources for historical Arabic manuscripts."
authors: ["Hakim Bouchal", "Ahror Belaid"]
year: 2023
venue: "arXiv:2312.07573"
source: "https://arxiv.org/abs/2312.07573"
doi: "10.48550/arxiv.2312.07573"
arxiv: "2312.07573"
licence: "CC BY-NC-ND 4.0 (arXiv)"
licence_url: "http://creativecommons.org/licenses/by-nc-nd/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/bouchal-2023-arabic-handwritten-text-line-dataset/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/bouchal-2023-arabic-handwritten-text-line-dataset/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2023](https://md.research.elamur.ai/free/en/papers/2023/page/7/) / Arabic Handwritten Text Line Dataset

# Arabic Handwritten Text Line Dataset

**Authors:** Hakim Bouchal, Ahror Belaid · **Institutions:** University of Béjaïa · **Year:** 2023 · **Venue:** arXiv:2312.07573 · **Licence:** [CC BY-NC-ND 4.0 (arXiv)](http://creativecommons.org/licenses/by-nc-nd/4.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2312.07573) · [PDF](https://arxiv.org/pdf/2312.07573) · [DOI 10.48550/arxiv.2312.07573](https://doi.org/10.48550/arxiv.2312.07573) · [arXiv 2312.07573](https://arxiv.org/abs/2312.07573) · [OpenAlex](https://openalex.org/W4389763515)

**Citations:** 1 citations (OpenAlex, 2026-10-14)

## Summary
The paper addresses a gap in resources for historical Arabic manuscripts. Line-level segmentation already has carefully annotated data, but the authors found no public dataset marking where individual words sit on the page. Word segmentation in old Arabic handwriting is hard because neighbouring words often touch or overlap, which defeats approaches that classify the gaps between connected components. The authors built KALIMA-Dataset on top of RASM2019, which provides polygonal line regions. They cut out the line images and then manually drew a bounding rectangle around every word in the main text. Each word received a transcription and its order within the line. The resulting corpus covers 497 lines from 26 pages of two scientific treatises. A t-SNE projection of the word images separates the two books into distinct clusters. The resource matters because it is the first word-level positional annotation for historical Arabic script, so it can support both word detection and word recognition research.

## Key points
- KALIMA-Dataset is derived from RASM2019, a collection of 120 historical scientific manuscript pages that originally had only line-level annotation.
- It contains 497 text-line images with 5062 bounding boxes. Of these, 4943 are Arabic words and 119 are other marks: 96 isolated letters and 23 Arabic numerals.
- The source is 26 pages from two British Library books, 13 pages each:
  - "Add MS 23494" is a music-theory treatise with 257 lines.
  - "Add MS 7474" is an astronomical work with 240 lines.
- Annotations come as two CSV files:
  - The first gives word coordinates relative to both the line and the page, with the transcription and the word's order in the line.
  - The second gives each line's position on the page.
- Marginal comments, geometric notation letters, symbols and numerals outside the main line are deliberately not treated as words. The data is hosted on GitHub (bouchalhakim/KALIMA_Dataset).

## Interesting ideas
- Instead of collecting new raw material, the authors added a finer annotation layer to an existing, trusted resource, which saves effort and keeps comparability with prior work. — For Mova (tree): New layers such as frame or coreference labels could be added on top of existing UD treebanks, reusing their sentence IDs so results stay comparable to standard parsing benchmarks.
- Non-word marks such as numerals, geometric letters and symbols are boxed but given their own categories, so they are tracked without polluting the word inventory. — For Mova (tokens): The tokenizer could give formula symbols, list markers and stray letters explicit non-word classes, so the tagger and lemmatizer never have to guess a POS for them.
- Each item is stored with coordinates in two frames of reference (line and page) plus its reading order, so it can be traced at any granularity. — For Mova (world): Entities and events in the world of a text could keep both sentence-local and document-level offsets with ordinal position, making every inferred fact traceable back to its exact source span.
- Annotation scope is limited to the main text line, and glosses that a line segmenter would likely cut off are excluded, so the labels match what the pipeline will actually see. — For Mova (learn): Silver data could exclude material that upstream steps systematically drop, such as headers or parenthetical asides, so training targets do not punish the parser for things it never receives.
- A quick t-SNE projection of the samples showed that the two source books form separate clusters, a cheap check of domain shift inside a dataset. — For Mova (learn): Projecting sentence or tree embeddings by source genre could reveal hidden domain splits in treebanks or silver data before training, and could guide how train and test splits are stratified.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/)

Topics: [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/)

## Cite

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@misc{bouchal2023arabic,
  title = {{Arabic Handwritten Text Line Dataset}},
  author = {Bouchal, Hakim and Belaid, Ahror},
  year = {2023},
  doi = {10.48550/arxiv.2312.07573},
  eprint = {2312.07573},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2312.07573}
}
```

- [← previous An ‘undervalued’ salvation oracle in the book of Jeremiah: Jeremiah 50:4–5 (Masoretic text)](https://md.research.elamur.ai/free/en/papers/terblanche-2023-undervalued-salvation-oracle-book-jeremiah/)
- [next → AraSpot: Arabic Spoken Command Spotting](https://md.research.elamur.ai/free/en/papers/salhab-2023-araspot-arabic-spoken-command-spotting/)
- [↑ back to papers of 2023](https://md.research.elamur.ai/free/en/papers/2023/page/7/)

## References

11 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. C. B. Rasmussen, K. Kirk, and T. B. Moeslund. The Challenge of Data Annotation in Deep Learning-A Case Study on Whole Plant Corn Silage. Sensors, vol. 22, no. 4, p. 1596, Feb. 2022, doi: 10.3390/s22041596. [↗](https://doi.org/10.3390/s22041596)
2. A. Al-Dmour and R. Zitar. Word Extraction from Arabic Handwritten Documents Based on Statistical Measures. Int. Rev. Comput. Softw., Jun. 2016. [find](https://scholar.google.com/scholar?q=A%2E+Al%2DDmour+and+R%2E+Zitar%2E+Word+Extraction+from+Arabic+Handwritten+Documents+Based+on+Statistical+Measures%2E+Int%2E+Rev%2E+Comput%2E+Softw%2E%2C+Jun%2E+2016%2E)
3. N. Aouadi and A. Echi. Word Extraction and Recognition in Arabic Handwritten Text. Int. J. Comput. Inf. Sci., vol. 12, pp. 17-23, Sep. 2016, doi: 10.21700/ijcis.2016.103. [↗](https://doi.org/10.21700/ijcis.2016.103)
4. J. H. AlKhateeb, J. Jiang, J. Ren, and S. S. Ipson. Component based Segmentation of Words from Handwritten Arabic Text. p. 5, 2009. [find](https://scholar.google.com/scholar?q=J%2E+H%2E+AlKhateeb%2C+J%2E+Jiang%2C+J%2E+Ren%2C+and+S%2E+S%2E+Ipson%2E+Component+based+Segmentation+of+Words+from+Handwritten+Arabic+Text%2E+p%2E+5%2C+2009%2E)
5. R. Ahmad, S. Naz, M. Z. Afzal, S. F. Rashid, M. Liwicki, and A. Dengel. KHATT: A Deep Learning Benchmark on Arabic Script. in 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), Kyoto, Nov. 2017, pp. 10-14. doi: 10.1109/ICDAR.2017.358. [↗](https://doi.org/10.1109/icdar.2017.358)
6. C. Clausner, A. Antonacopoulos, N. Mcgregor, and D. Wilson-Nunn. ICFHR 2018 Competition on Recognition of Historical Arabic Scientific Manuscripts - RASM2018. in 2018 16th International Conference on Frontiers in Handwriting Recognition (ICFHR), Niagara Falls, NY, USA, Aug. 2018, pp. 471-476. doi: 10.1109/ICFHR-2018.2018.00088. [↗](https://doi.org/10.1109/icfhr-2018.2018.00088)
7. Reza Farrahi Moghaddam, Mohamed Cheriet, Mathias M. Adankon, Kostyantyn Filonenko, and Robert Wisnovsky. IBN SINA: A database for research on processing and understanding of Arabic manuscripts images. Proceedings of DAS'10, June 9-11, 2010, Boston, MA, USA [find](https://scholar.google.com/scholar?q=Reza+Farrahi+Moghaddam%2C+Mohamed+Cheriet%2C+Mathias+M%2E+Adankon%2C+Kostyantyn+Filonenko%2C+and+Robert+Wisnovsky%2E+IBN+SINA%3A+A+database+for+research+on+processing+and+understanding+of+Arabic+manuscripts+images%2E)
8. M. Kassis, A. Abdalhaleem, A. Droby, R. Alaasam and J. El-Sana, VML-HD: The historical Arabic documents dataset for recognition systems. 2017 1st International Workshop on Arabic Script Analysis and Recognition (ASAR). 2017, pp. 11-14, doi: 10.1109/ASAR.2017.8067751. [↗](https://doi.org/10.1109/asar.2017.8067751)
9. A. Shahkolaei, A. Beghdadi, S. Al-maadeed and M. Cheriet. MHDID: A Multi-distortion Historical Document Image Database. 2018 IEEE 2nd International Workshop on Arabic and Derived Script Analysis and Recognition (ASAR), 2018, pp. 156-160, doi: 10.1109/ASAR.2018.8480372 [↗](https://doi.org/10.1109/asar.2018.8480372)
10. Benjamin Kiessling, Daniel Stökl Ben Ezra, Matthew Thomas Miller. BADAM: A Public Dataset for Baseline Detection in Arabic-script Manuscripts. In: Proceedings of the 5th International Workshop on Historical Document Imaging and Processing (HIP). pp. 13-18 (2019) [find](https://scholar.google.com/scholar?q=Benjamin+Kiessling%2C+Daniel+Sto%CC%88kl+Ben+Ezra%2C+Matthew+Thomas+Miller%2E+BADAM%3A+A+Public+Dataset+for+Baseline+Detection+in+Arabic%2Dscript+Manuscripts%2E+In%3A+Proceedings+of+the+5th+International+Workshop+on+His)
11. Melit Devassy B, George S. Dimensionality reduction and visualisation of hyperspectral ink data using t-SNE. Forensic Sci Int (2020) 311:110194. doi: 10.1016/j.forsciint.2020.110194 [↗](https://doi.org/10.1016/j.forsciint.2020.110194)