Life sciences · Preprint

EMBLEM: Enhancing Multi-script Table Detection through Masking

arXiv · September 8, 2026

Preliminary Why this matters to you Expert discussion

Early or partial results. Treat as a signal, not a conclusion.

AI Summary

EMBLEM is a masking-based framework and MANDALA is a new multilingual dataset for table detection in documents spanning 18 languages and 15 scripts. The method achieves a 20.8% absolute F1-score gain on the new dataset by training on masked English images and applying script-agnostic layout learning. This is an unreviewed technical contribution without validation of downstream utility or clinical relevance.

What was studied

Algorithmic development and comparative evaluation on a new dataset. 2,323 table-containing document pages in 18 languages and 15 scripts across diverse domains; no human subjects.. Intervention: EMBLEM masking-based paradigm for multi-script table detection applied to masked images. Compared with: Strong baselines; five standard English-dominant benchmarks. n = 2,323.

Key findings

MANDALA dataset contains 2,323 manually curated table-containing pages spanning 18 languages and 15 scripts across diverse domains EMBLEM achieves an absolute F1-score gain of 20.8% on MANDALA using only English masked images for fine-tuning with no multi-script training data EMBLEM remains competitive on five standard English-dominant benchmarks while outperforming strong baselines on MANDALA

Safety

Safety was not reported in the material analysed. Check the source before drawing any conclusion about harm.

Who this may help

The source did not state who this applies to in practice.

Plain-language summary

A methodological contribution with a new dataset and algorithm for table detection, but this is a preprint computer science work without peer review, clinical outcomes, or human validation studies.

Study design

As stated by the source record.

MethodsA manually curated dataset (MANDALA) of 2,323 table-containing pages was created. EMBLEM, a masking-based approach, was applied to generate masked images that conceal script- and font-specific details. Models pre-trained on English documents were fine-tuned using only English masked images. Performance was evaluated across three table detection architectures on MANDALA and on five existing English-dominant benchmarks using F1-score as the metric.

DesignAlgorithmic development and comparative evaluation on a new dataset

Population2,323 table-containing document pages in 18 languages and 15 scripts across diverse domains; no human subjects.

Sample size2,323

InterventionEMBLEM masking-based paradigm for multi-script table detection applied to masked images

ComparatorStrong baselines; five standard English-dominant benchmarks

Reported figures

Quoted from the source exactly as published.

MANDALA dataset: pages2,323

MANDALA dataset: languages18

MANDALA dataset: scripts15

Absolute F1-score gain on MANDALA (masked English images, no multi-script training)20.8%

Table detection architectures tested3

Caveats

Preprint has not undergone peer review
No statistical significance testing, confidence intervals, or p-values reported
F1-score is a technical metric; real-world downstream impact (information retrieval, document reconstruction, VQA) not measured
Generalization to unlisted languages, scripts, or document types not explicitly demonstrated

Evidence assessment

Graded across the dimensions that decide whether you should act, each from what the source actually supports. There is no single score, and where a dimension was not assessed it says so.

Study design

limited

Algorithmic comparison across architectures; no randomization, blinding, or control group structure.

Directness

limited

Evaluates technical task (table detection) on a curated dataset; no direct measurement of clinical, patient, or real-world downstream impact.

Consistency

moderate

EMBLEM tested on three architectures; performance consistent across them and competitive on five existing benchmarks.

Risk of bias

insufficient

No mention of blinding, pre-registration, or conflict-of-interest disclosure in preprint abstract.

Sample adequacy

moderate

MANDALA dataset contains 2,323 table-containing pages across 18 languages and 15 scripts; size not justified for statistical power.

Source abstract

Table detection is a core task in document analysis, supporting downstream applications such as information retrieval, document reconstruction, and visual question answering. While existing deep learning models perform well on English and Chinese documents, they struggle with multilingual, multi-script documents due to script diversity and the limited availability of labeled data. To address this challenge, we introduce MANDALA (Multi-script Annotated Documents for Table Detection), a manually curated dataset of 2,323 table-containing pages spanning 18 languages and 15 scripts across diverse domains. We also propose EMBLEM, a masking-based paradigm for Multi-script Table Detection (MTD). EMBLEM generates masked images that conceal script- and font-specific details, enabling models pre-trained on abundant English documents to focus on script-agnostic page layout. Experiments across three table detection architectures show that EMBLEM consistently outperforms strong baselines on MANDALA while remaining competitive on five standard English-dominant benchmarks. Using only English masked images for fine-tuning, with no multi-script training data, EMBLEM achieves an absolute F1-score gain of 20.8% on MANDALA. We release MANDALA along with the accompanying code and models at https://github.com/IITB-LEAP-OCR/EMBLEM.git.

Follow from this

Taken from the source record, never inferred. Follow any of these and new work involving them reaches your briefing.

Ganesh RamakrishnanUdhay BrahmiDhruv Kudale