LV-ROVER-MLT Low-Resource Maltese OCR by Multi-Stream Voting

Jul 16, 2026

Upload by Finn

Maltese OCR is hindered by limited labelled data and a diacritic-rich Latin script with silent font substitutions and hyphen conventions that depend on form rather than lexicon alone. To address scarce paragraph-level training material, a synthetic training pipeline was built and a 5-stream Tesseract ensemble was implemented using a lexicon-anchored, ROVER-style voting scheme adapted to low-resource conditions. Benchmark results on 422 paragraphs show improved character error rates: 0.0234 baseline, 44% better with ensemble voting, and 0.00700 after label normalization and diacritic recovery, with related tests on Hungarian and Luxembourgish.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages11

Reading Time28 min

Abstract

Introduction

Key challenges

System overview

Evaluation setup and contributions

LV-ROVER-MLT: Low-Resource Maltese OCR by Multi-Stream Voting Adam Darmanin Independent Researcher [adamdarmanin@hecatusresearch.com](adamdarmanin@hecatusresearch.com) arXiv :2607 .00250v2 [ cs .CL] 2 Jul 2026 Abstract Maltese although a low resource language, has its own text corpora and pretrained language models [20], but we are aware of only one real labelled PDF corpus for OCR training, 57 pages [25], far below what paragraph-level training needs. With no real corpus to train on at scale, we built a synthetic training pipeline and a 5-stream Tesseract ensemble voted under a lexicon-anchored, ROVER-style scheme adapted for a lowresource setting [10, 24] . We call the Maltese submission LVROVER-MLT: an engineered adaptation of LV-ROVER’s voting algorithm, not a new one, submitted to the DocEng 2026 competition. All results below are dev-set figures from the competition’s own benchmark; the held-out real test CER is unknown at the time of writing and this paper does not claim one. We report results on a 422-paragraph benchmark against a fine-tuned-Tesseract baseline of character error rate (CER) 0.0234. Ensemble recognition alone, scored under the same label convention as the baseline, improves CER by 44 percent to 0.01317 . A post-processing chain that aligns Tesseract’s straight-quote and dash output to the benchmark’s curlyquote convention, plus one stage that recovers misread diacritics, brings the full pipeline to CER 0.00700, a 70 percent reduction. We also tested the same method, unchanged, on Hungarian and Luxembourgish: a bootstrap and permutation audit confirms a 33.7 percent CER improvement on Luxembourgish, while the Hungarian margin (0.8 percent) is not statistically significant. Keywords OCR, low-resource languages, Maltese, synthetic data, character error rate Working paper, not peer reviewed. 1 Introduction Maltese is a Semitic language written in Latin script, spoken by approximately half a million native speakers. Centuries of archival text in Maltese remain unsearchable because they have not been digitised at scale, and OCR is the prerequisite step. The script itself is a diﬀicult OCR target: its 30-letter alphabet extends the standard Latin inventory with Ċċ, Ġ ġ, Ħ ħ, Ż ż, and the digraphs Għ għ and Ie ie, fonts frequently silently substitute the base letter for the diacritic form, and the definite article attaches to the following noun via a structural hyphen (il-kelb, id-dar, fis-se􀀀􀀀) that shares the same glyph, -, with soft line-break hyphens. The only labelled real Maltese PDF data we are aware of comes from the NOMOCRAT annotation project [25]: 57 PDF pages, line- and paragraph-transcribed per the project’s own description, which we did not independently re-derive. That is far below what training a paragraph-level recogniser needs, and we found no indication it has been released for reuse outside that project. Three technical challenges faced by this system. First, any tokeniser or font that silently substitutes c for ċ corrupts labels at the encoder or rendering stage; we treat the four diacritic pairs (ċ/c, ġ/g, ħ/h, ż/z) as canaries, tracked at every stage of the pipeline. Second, the soft-versus-structural hyphen distinction is not lexically decidable on the glyph alone: both surfaces use-, so resolution requires either a language model or a rulebased joiner with Maltese morphological knowledge [6, 18] . Third, the gold label convention for a benchmark shapes which improvements look real. The dev-gold labels use curly quotes ( ‘ ’ “ ”) and an em-dash (—); Tesseract’s raw output uses straight ASCII. Normalising one to the other produces a large CER drop that has nothing to do with recognition quality. We make this explicit with a dual-CER protocol. This paper describes the LV-ROVER-MLT system: five parallel Tesseract LSTM streams voted per word under a soft Maltese lexicon, followed by a five-stage label normalisation chain and a rule-based line joiner. We evaluate

FAQs

What problem does LV-ROVER-MLT target in Maltese OCR?

It targets low-resource Maltese OCR where labelled paragraph-level training data is scarce and where diacritics and hyphen conventions make text recognition error-prone.

How does the system improve results beyond a single Tesseract model?

It uses a 5-stream Tesseract ensemble with lexicon-anchored, ROVER-style multi-stream voting, followed by label normalization and rule-based line joining with diacritic recovery.

Why are separate CER improvements reported?

The paper uses a dual-CER protocol to distinguish recognition quality gains from drops caused by differences in label conventions (straight quotes vs curly quotes, and ASCII dash vs em-dash).

Jul 16, 2026

Upload by Finn