CoChat

The DOI matches its Crossref record; 30 of the 30 references it cites were found in a registry or an index; none of the 11 sources its claims cite is checked for support yet.

A Dataset of Wolof Ajami Manuscripts for HTR and OCR

Open access
Yousuf, Oreen, Diagne, Elhadji Djibril, Høgel, Christian, Megyesi, Beata, Nivre, Joakim · 2026 · Proceedings of the Language Resources and Evaluation Conference

Abstract

We present the first ever dataset of manually segmented and transcribed Ajami manuscripts written in Wolof.The term Ajami refers to modified Arabic-script orthographies used to transcribe African languages.Handwritten text recognition (HTR) and optical character recognition (OCR) models for Arabic-script languages perform poorly on African languages written in Ajami orthographies because these languages are not represented in the pre-training data of the models.This leads to recognition models being unable to extract unique Arabic-script letters and ubiquitous diacritics used in African languages, and struggling to adapt to various calligraphy styles used across Africa.We release the following as an open-source dataset: an ALTO formatting of high-quality images of handwritten and printed, 20th-century Wolof manuscripts; manual segmentation (region and line); and manual transcriptions.We extend our contribution by evaluating several Arabic-script recognition models intended for historical manuscripts and find they produce character error rates (CER) of 61-81%.Transcriptions produced by the evaluated recognition models, as well as a keyboard to transcribe Wolof Ajami manuscripts, are released as well.The digitally transcribed text in the dataset can also be utilized for various natural language processing (NLP) and historical linguistic tasks.

Overview

CoChat Summary

TL;DR

The paper presents the first dataset of manually segmented and transcribed Ajami manuscripts written in Wolof, comprising 19 20th-century manuscripts (4,743 lines, 313 pages) in ALTO format plus a specialized keyboard. It evaluates OpenITI Arabic-script recognition models and reports high character error rates of 61–81%.

Key findings

  • The dataset contains 19 20th-century Wolof Ajami manuscripts covering 4,743 lines and 313 pages; 9 prose and 10 poetry manuscripts; 8 handwritten and 11 printed; all in Maghrebi calligraphy; provenance Touba, Senegal.
  • Evaluated OpenITI Arabic-script HTR/OCR models produce CERs of 61–81% on Wolof Ajami; MS Mellon Print performs best (CER 0.625) and Urdu Print worst (CER 0.822).
  • Models struggle to recognize Arabic-script diacritics (tashkīl) and unique Wolof Ajami letters/diacritics absent from their pre-training languages; top three models better capture general line structure but omit diacritics and unique Wolof characters.
  • The authors release the ALTO-formatted dataset, model transcriptions, and a Keyman keyboard for transcribing historical Wolof Ajami manuscripts as open source.

Methods

  • Curated 19 20th-century Wolof Ajami manuscripts from Boston University’s Ajami collection (“The Four Languages” and “African Ajami Library”); dataset includes high-quality images, manual region and line segmentation, line order, and manual transcriptions in ALTO format.
  • Annotation used eScriptorium; each manuscript was transcribed by a philologist with Wolof Ajami expertise and then independently reviewed for error correction.
  • Baseline experiments evaluated OpenITI’s suite of Arabic-script recognition models for historical manuscripts (MS Mellon Print, MS Pretrained, Generalized, Mellon Ottoturk Print, Mellon Print, Pretrained Print, Urdu Print) using character error rate (CER).

Limitations

  • The paper does not include a dedicated limitations section; it states that existing recognition models perform poorly due to lack of appropriate training data for Wolof Ajami orthography and the heavy use of diacritics and unique letters absent from pre-training languages.
  • Evident from the text: the dataset is limited to 20th-century Wolof Ajami manuscripts from Touba, Senegal, all in Maghrebi calligraphy, sourced from Boston University’s Ajami collection (8 handwritten, 11 printed).
  • Evident from the text: the manuscript count is inconsistent—Section 2 reports 19 manuscripts, while the Conclusion reports 17 manuscripts; baseline evaluation is limited to OpenITI’s suite of Arabic-script models.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.63317/4pz98ojeeqpw

Cites

30 of 30 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

0 Valid · 0 Not valid · 7 Not checked · 33 Own results and interpretations

18 more claims

Create a free account to see all 40 claims and to verify the 2 references still open.

See all claims — free account
  • It may state two separate claims as one. 13
  • Probably not a substantial claim. 5

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Research works printed but not linked: 2

  1. Maxim Romanov, Matthew Thomas Miller, Sarah Bowen Savant, and Masoumeh Seydi. 2019. Open islamicate texts initiative: A machine-readable corpus of texts produced in the premodern islamicate world (poster). In Digital Humanities 2019 Conference Papers (9-12 July 2019). Utrecht University, 2019.
  2. Eddie Antonio Santos and Atticus Harrigan. 2020. Design and evaluation of a smartphone key- board for plains cree syllabics. In Proceedings of the 1st Joint Workshop on Spoken Language Technologies for Under-Resourced Languages (SLTU) and Collaboration and Computing for Under-Resourced Languages (CCURL), pages 88-96.

Metadata

Authors
Yousuf, O., Diagne, E. D., Høgel, C., Megyesi, B., & Nivre, J.
Affiliations
  • Uppsala University (SE)
  • Lund University (SE)
Venue
Proceedings of the Language Resources and Evaluation Conference
Published
2026-05
Pages
3234-3239
ISSN
2522-2686
OpenAlex
W7160193388
Cited by
0
References
9
FWCI
1.29
Type
conference paper
Language
en
Open access
Locations
Language Resources and Evaluation Conference (LREC) (publishedVersion)
Funders
  • Vetenskapsrådet (2022-06343)
  • Riksbankens Jubileumsfond (M24-0028)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Handwritten Text Recognition Techniques
  • Social Sciences › Arts and Humanities › Literature and Literary Theory › Digital Humanities and Scholarship
  • Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
Keywords
Ajami script, Wolof language, handwriting recognition, optical character recognition, historical manuscripts, character error rate, diacritics, calligraphic style, transcription, low-resource languages

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-16, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-16 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.