ACADEMIA & RESEARCH
Research papers translated without breaking a single equation
Ellon AI translates LaTeX papers, theses, and journal submissions while preserving mathematical notation, citation syntax, and bibliography formatting. for international collaboration and multi-language conference submissions.
Academic translation shouldn't wreck the math
A researcher submitting a paper to an international journal, preparing a thesis defense in a second language, or collaborating with a lab overseas runs into the same problem: traditional translation tools treat LaTeX like plain text. Equation environments get scrambled. BibTeX citations get corrupted. Greek letters end up translated to English words. The researcher's paper, which compiled perfectly before translation, now produces dozens of compilation errors on the translated version. The usual workaround. translate the .tex source carefully by hand, then spend hours fixing what the tool broke. defeats the point of having a translation tool. Ellon AI treats LaTeX as structured content: equations, citations, and math environments pass through unchanged; only the prose text translates.
- LaTeX equations, math environments, and Greek letters preserved
- BibTeX citations and bibliography references pass through unchanged
- Cross-references, labels, and hyperref links intact
- Markdown academic formats (.md) fully supported
Translate a paper without recompiling the math
Upload a .tex source, Markdown draft, or compiled PDF. Ellon AI translates the prose while equations, inline math with Greek subscripts, figure and table references, BibTeX citation calls, author metadata, and the bibliography block all pass through unchanged. The translated source compiles identically to the original. title, abstract, numbered sections, and Table 1 all line up with the source manuscript, just in the target language.
Original · Français
Nous présentons une approche hybride combinant l’apprentissage contrastif et la régularisation de cohérence pour la classification d’images médicales en régime de faible étiquetage. Notre méthode atteint une précision de 94,3 % sur le jeu de données CheXpert avec seulement 10 % d’étiquettes, surpassant l’état de l’art antérieur de 2,8 points de pourcentage. Le code source et les hyperparamètres sont disponibles dans [2].
Mots-clés — apprentissage semi-supervisé ; classification d’images médicales ; apprentissage contrastif ; régularisation de cohérence
1. Introduction
L’annotation manuelle d’images médicales requiert une expertise clinique et demeure coûteuse à grande échelle [1]. Pour pallier cette rareté, les approches semi-supervisées exploitent les données non étiquetées lors de l’apprentissage [3, 4]. Cependant, les méthodes existantes souffrent d’une instabilité lors du passage d’un domaine radiologique à un autre [5]. Dans le présent travail, nous proposons un cadre hybride qui combine l’apprentissage contrastif, la régularisation de cohérence et une pondération adaptative.
où ℒ_sup, ℒ_con et ℒ_reg désignent respectivement les pertes supervisée, contrastive et de régularisation.
2. Méthode
Notre modèle combine un encodeur partagé f_θ avec deux têtes de projection g_ϕ et h_ψ (cf. Fig. 1). La perte totale s’écrit comme la somme pondérée de l’équation (1), où λ et μ sont ajustés dynamiquement par planification cosinus selon [6]. Le réseau est entraîné pendant 200 époques avec un taux d’apprentissage initial de 1,0 × 10⁻³ et un lot de 128 images.
Fig. 1 — Architecture du modèle à double projection. L’encodeur partagé f_θ alimente les têtes contrastive g_ϕ et de cohérence h_ψ.
3. Résultats
Comme le montre le tableau 1, notre méthode surpasse les modèles de référence sur les trois jeux de données évalués. Le gain le plus important est observé sur CheXpert en régime de 10 % d’étiquettes (+2,8 points). L’étude d’ablation confirme que la régularisation ℒ_reg contribue pour 1,3 point, le reste provenant de la planification cosinus.
| Jeu de données | État de l’art | Ours |
|---|---|---|
| CheXpert (10 %) | 91,5 (0,3) | 94,3 (0,2) |
| ISIC-2020 | 87,2 (0,5) | 89,1 (0,4) |
| HAM-10000 | 83,7 (0,6) | 85,4 (0,5) |
Tableau 1 — Précision (%) sur trois jeux de données. Moyenne de 5 exécutions ; écart-type entre parenthèses.
Références
- [1]Litjens, G. et al. (2017). A survey on deep learning in medical image analysis. Medical Image Analysis, 42, 60–88.
- [2]Dubois, M., Roussel, J.-L. (2026). Code and hyperparameters for the hybrid semi-supervised classifier. Zenodo. https://doi.org/10.5281/zenodo.XXXXXX
- [3]Sohn, K. et al. (2020). FixMatch: Simplifying semi-supervised learning with consistency and confidence. NeurIPS.
Translated · English
We present a hybrid approach combining contrastive learning and consistency regularization for medical image classification in low-label regimes. Our method achieves 94.3% accuracy on the CheXpert dataset with only 10% of labels, outperforming the prior state of the art by 2.8 percentage points. Source code and hyperparameters are available at [2].
Keywords: semi-supervised learning; medical image classification; contrastive learning; consistency regularization
1. Introduction
Manual annotation of medical images requires clinical expertise and remains costly at scale [1]. To address this scarcity, semi-supervised approaches leverage unlabeled data during training [3, 4]. However, existing methods suffer from instability when moving across radiological domains [5]. In this work, we propose a hybrid framework that combines contrastive learning, consistency regularization, and an adaptive weighting schedule.
where ℒ_sup, ℒ_con, and ℒ_reg denote the supervised, contrastive, and regularization losses, respectively.
2. Method
Our model combines a shared encoder f_θ with two projection heads g_ϕ and h_ψ (see Fig. 1). The total loss is the weighted sum in Equation (1), where λ and μ are tuned dynamically by cosine scheduling per [6]. The network is trained for 200 epochs with an initial learning rate of 1.0 × 10⁻³ and a batch size of 128 images.
Fig. 1 — Dual-projection model architecture. Shared encoder f_θ feeds the contrastive head g_ϕ and the consistency head h_ψ.
3. Results
As shown in Table 1, our method outperforms baseline models across all three evaluated datasets. The largest gain is observed on CheXpert in the 10%-label regime (+2.8 points). The ablation study confirms that the ℒ_reg regularization contributes 1.3 points, with the remainder from the cosine schedule.
| Dataset | Prior SOTA | Ours |
|---|---|---|
| CheXpert (10%) | 91.5 (0.3) | 94.3 (0.2) |
| ISIC-2020 | 87.2 (0.5) | 89.1 (0.4) |
| HAM-10000 | 83.7 (0.6) | 85.4 (0.5) |
Table 1 — Accuracy (%) across three datasets. Mean over 5 runs; standard deviation in parentheses.
References
- [1]Litjens, G. et al. (2017). A survey on deep learning in medical image analysis. Medical Image Analysis, 42, 60–88.
- [2]Dubois, M., Roussel, J.-L. (2026). Code and hyperparameters for the hybrid semi-supervised classifier. Zenodo. https://doi.org/10.5281/zenodo.XXXXXX
- [3]Sohn, K. et al. (2020). FixMatch: Simplifying semi-supervised learning with consistency and confidence. NeurIPS.
Audit every revision across a response-to-reviewers
LaTeX has no native tracked-changes primitive, so academics typically maintain revision rounds in a DOCX response-to-reviewers document. Upload the old and new response (DOCX or PDF) and Ellon AI produces a Word tracked-changes file covering every reviewer point. updated quantitative claims, methodology clarifications, added citations, regenerated tables, new limitations paragraphs. plus a semantic summary separating substantive scientific edits from formatting polish. Designed for the peer-review cycle, from R1 through acceptance.
We thank the reviewers for their careful reading and constructive suggestions. This R2 submission addresses every numbered point in the editor’s letter of 5 March 2026. Revised text is shown inline with tracked changes; new citations are marked “new ref.” and appear at the end of the reference list. Unchanged sections carry over from R1 unmodified.
Our method achieves an accuracy of 92.8% an accuracy of 94.3% (± 0.2) on the CheXpert dataset with only 10% of labels, outperforming the prior state of the art by more than two percentage points 2.8 percentage points (p < 0.001, paired t-test). Source code and hyperparameters are available at [2]; the trained model weights will be released at acceptance.
The total loss is the weighted sum in Equation (1), where λ and μ are set to a constant value of 0.5 tuned dynamically via cosine scheduling following the protocol in [6] — rising from 0 to 0.75 over the first 50 epochs and held thereafter. The revised scheduling produces a 0.6-point gain over the constant-weight ablation (Appendix C).
We have added a full discussion of class-imbalance handling in §3.2 (new in R2). For CheXpert, we use the 5-class subset rather than the 14- class taxonomy; the class distribution is reported in Table 3 and reflects the natural incidence in the ambulatory screening cohort. Minority classes are weighted inversely to their frequency during the supervised term, with the contrastive and consistency terms remaining unweighted.
All reported accuracy values are now means over three five independent runs; standard deviations appear in parentheses in Table 1. Statistical significance was tested with the Wilcoxon signed-rank test on per-run accuracies across the five seeds, which produced p < 0.001 against the strongest prior method on CheXpert and ISIC-2020, and p = 0.012 on HAM-10000.
We have incorporated additional prior work the reviewer identified: Zhang et al. (2024) on label-efficient contrastive representations [16, new ref.], Kim & Park (2023) on cosine scheduling for SSL [17, new ref.], and Martinez et al. (2025) on domain-shifted evaluation protocols [18, new ref.]. The revised Related Work section (§2) now properly situates our contribution with respect to these.
Table 1 has been regenerated with three five-seed means and standard deviations. Figure 3 is new in R2: a learning-curve plot comparing our method against the strongest baseline across the three datasets, confirming that the convergence behaviour is stable and does not depend on lucky initialization.
The conclusions section now includes an explicit limitations paragraph (§5.3), covering: (i) the dependence of our cosine schedule on domain-specific tuning in the first 30 epochs; (ii) the degraded performance observed when the labeled-to- unlabeled ratio exceeds 1:50; and (iii) the absence of external validation on a held-out hospital cohort, which we flag as the most important direction for future work.
A full copy-edit pass was performed by the corresponding author before resubmission. Figure captions, Equation numbering, and BibTeX entries are consistent across the revised manuscript and the supplementary material.
How researchers use Ellon AI
Academic work is increasingly international. PhD candidates publish in English regardless of their program's language. Research labs collaborate across continents. Conference submissions go to venues that mix English, French, German, Spanish, Japanese, and Chinese. The translation workload this creates has historically fallen on junior researchers who spent days or weeks translating their own work or their lab's work. time that should have been spent on research.
Journal submission and peer review
For researchers writing in a second language, the final polish before submission to an English-language journal is often the most grueling step. Ellon AI translates the French, Spanish, German, or Chinese draft into publication-grade English while preserving LaTeX math, citation syntax, and figure references. The researcher reviews the output for nuance and tone rather than translating the entire paper from scratch. Peer review responses and revised submissions go through the same workflow for subsequent rounds.
Thesis and dissertation
PhD candidates at international programs often need to produce their thesis in the program's official language, in English for the broader academic audience, or both. Ellon AI handles thesis-length documents with chapter structure, footnote numbering, and bibliography references intact. The compare tool is useful during the review process to track advisor feedback across revisions. semantic analysis separates substantive advisor comments from minor edits.
Conference submissions and proceedings
Academic conferences typically require submissions in English, even when the research community is primarily non-English-speaking. Ellon AI lets researchers prepare English-language submissions quickly from non-English drafts, respecting the LaTeX templates conferences distribute (IEEE, ACM, Springer LNCS).
International collaboration
Co-authored papers across labs in different countries produce drafts, comments, and revisions in multiple languages. Ellon AI handles the translation across drafts so the collaboration moves at the speed of the research, not the speed of manual translation. For recurring collaborations, terminology consistency is maintained within each translated document.
Literature review and meta-analysis
Systematic reviews often require reading primary sources across multiple languages. a meta-analysis in clinical research might pull from English, German, Japanese, and Chinese primary studies. Ellon AI handles the bulk translation of foreign-language papers so the research team can focus on synthesis rather than language translation.
Grant applications and research funding
Grant applications to international funding bodies. European Research Council, NSF, JSPS, DFG. often require specific language versions. Ellon AI translates grant applications, research statements, and project narratives while preserving the structural formatting funders expect (budget tables, timeline diagrams, reference lists).
Teaching materials and syllabi
For instructors teaching at international programs or running exchange programs, course materials, lecture notes, problem sets, and exams translate across the languages students operate in. Ellon AI handles the technical terminology consistency (physics notation, biological nomenclature, chemical formulas) that teaching materials depend on.
Scholarly communication and outreach
Research communication. summary articles for non-specialist audiences, press releases about findings. translates for international outreach while preserving figures and data references.
Revision cycles
Journal peer review produces major revisions, minor revisions, and responses to reviewer comments. When those revisions happen across languages (a Spanish-language research group responding in English to reviewers at a US journal), the compare tool surfaces what actually changed between revision rounds with AI semantic analysis. Co-authors approve the changes in minutes instead of re-reading the full paper.
Format coverage
LaTeX (.tex with standard academic classes), Markdown, Word manuscripts, and PDF drafts are all supported. Output compiles or renders in the target format identically to the source.
Language pairs
Ellon AI supports 200+ language pairs. Common academic pairs. English to/from French, German, Spanish, Portuguese, Italian, Japanese, Korean, Mandarin, Russian. handle discipline-specific terminology across STEM, humanities, and social sciences with consistency appropriate for peer review.
Cost considerations
Per-page academic translation pricing through traditional translation services historically ran into hundreds of dollars per thousand words, making comprehensive translation impractical for junior researchers. Ellon AI's per-page pricing makes ongoing translation workflows. for grant cycles, journal submissions, and ongoing collaborations. economically viable at the scale actual research programs operate on.
The practical outcome: language stops being a research bottleneck. Researchers publish, collaborate, and communicate in the language the audience reads, on the same timeline as the research itself.
Frequently asked
Frequently asked questions
Still have questions?
Reach out and we'll answer within one business day.
Submit your next paper in the journal's language
Start with 20 free pages. No credit card. Upload a LaTeX paper and see the equations and citations survive the translation.
Start your free trial