We describe the Charles University CUNI-AR submission to the WMT26 General Translation shared task for English into Egyptian Arabic. Our pipeline is built around GEMBA (LLM-as-judge quality) estimator that scores translations independently for accuracy and dialect authenticity. We use it throughout our pipeline: to filter training data, to construct preference pairs for DPO, to select best checkpoints, and to pick the best hypothesis per segment at inference time. Our primary translation model is Gemma-4-12B-it finetuned via QLoRA, we also finetune Jais-2-8B-Chat and Aya-expanse-8B. Gemma-4-12B-it provides a high-quality Egyptian Arabic translation, so SFT yields only marginal quality gains in automatic metrics; DPO is able to improve the metrics it optimizes for, but the real gain in translation quality is to be assessed by human evaluation. The final submission selects the best-scoring translation per segment from 31 candidate checkpoints, with the unmodified base model contributing 31\% of the output.

We describe our constrained CUNI-MH-v3 (based on Qwen3-14B) and unconstrained CUNI-EdUKate (based on Granite 4.1 8B) submissions for the WMT26 General MT shared task. Both submissions use a three-stage training procedure: QLoRA supervised fine-tuning on synthetic sentence pairs, continued supervised fine-tuning on a smaller mixture of paragraph-level and instruction-oriented examples, and iterative preference optimization with automatically generated and scored preference pairs. We document the data construction, training setup, checkpoint selection, and evaluation. We also share the training data and model checkpoints for our constrained system.

Text-in-image translation aims to translate all textual content embedded in an image. In Text Image Translation (TIT), the output is translated text; In-Image Machine Translation (IIMT) additionally renders the translation back into the image. We introduce ImgMT, a massively multilingual dataset and benchmark mined from localized SVG illustrations on Wikimedia Commons, focusing on diagrams, maps, and infographics. ImgMT contains 35,612 aligned image pairs grouped into 1,855 image sets, spanning 129 languages, 20+ writing systems, and 4,886 language pairs.

We propose a multi-level evaluation protocol that measures source-side text-region detection and OCR accuracy alongside target-side translation quality. We benchmark several multimodal large language models and a dedicated TIT system in English-to-X and X-to-English settings, and find that the source language and writing system substantially affect TIT quality. Finally, we show that ImgMT is not well suited for isolating the contribution of visual information in multimodal translation.

We present CzechDocs, a multiway parallel dataset of formatted documents (HTML, DOCX, and PDF) covering Czech and minority languages used in Czechia—primarily Ukrainian and English, with smaller portions of Vietnamese, Russian and other languages. The dataset is designed to support the evaluation of machine translation systems that aim to preserve document formatting during translation. We provide a comparison of the most common approaches to format-preserving machine translation on a validation subset of the dataset. This validation split, together with the evaluation toolkit, is publicly released for further research. A held-out test split will be reserved for a future shared task focused on document-level translation with formatting preservation.

This thesis argues that the currently widely used Natural Language Processing algorithms possibly have various limitations related to the properties of the texts they handle and produce. With the wide adoption of these tools in rapid progress, we must ask what these limitations are and what are the possible implications of integrating such tools into our daily lives.As a testbed, we have chosen the task of Neural Machine Translation (NMT). Nevertheless, we aim for general insights and outcomes, applicable to current Large Language Models (LLMs). We ask whether the algorithms used in NMT have inherent inductive biases that are beneficial for most types of inputs but might harm the processing of untypical texts, thereby contributing to a cycle of monotonous, repetitive language – whether generated by machines or humans.To explore this hypothesis, we define a set of measures to quantify text diversity based on its statistical properties, like uniformity or rhythmicity of word-level surprisal, on multiple scales (sentence, discourse, language). We conduct a series of experiments to investigate whether NMT systems struggle with maintaining the diversity of such texts, potentially reducing the richness of the generated language, compared to human translators.We further analyze potential origins of these limitations within existing training objectives and decoding strategies. Ultimately, our goal is to propose and validate alternative approaches (e.g., loss functions, decoding algorithms) that maintain the diversity and complexity of language and that allow for better global planning of the output generation, enabling the models to better reflect the ambiguities inherent in human communication.

We describe the Charles University (CUNI) submission to the AlexandriaX-2026 Subtask 1, a shared task on English-to-Arabic dialectal conversation translation covering 13 Arabic varieties. Our approach fine-tunes Gemma-4-31B-it with QLoRA on the Alexandria training data at the conversation level, using rich contextual prompts that include domain metadata, participant names and speaker labels. We explore multiple training data mixes and hyperparameters, yielding a pool of candidate translations for the test set. These are reranked with a genetic-algorithm-enhanced minimum Bayes risk (GA-MBR) decoding using sentence-level spBLEU as the fitness function. Fine-tuning improves over the unmodified base model by approximately 5.2 ChrF points on the development set (from 45.2 to 50.4 ChrF for the best single checkpoint, or 25.1 to 32.6 spBLEU); GA-MBR reranking over a 40-candidate development pool yields 51.3 ChrF / 34.0 spBLEU, a further improvement of 0.9 ChrF points over the best pool member. Our submission ranked second in the unconstrained track.

This work presents an evaluation of large language models (LLMs) for English to dialectal Arabic machine translation on the MADAR dataset. We evaluate both translation directions (English to Arabic and vice-versa) on 16 Arabic dialects. Our experiments cover a diverse set of models, including specialized Arabic models (Jais, Nile), multilingual models (Gemma, Command-R, Mistral, Aya), and commercial APIs (GPT-4.1). We employ multiple evaluation metrics: BLEU, CHRF, COMET (both reference-based and reference-less variants) and GEMBA (LLM-as-a-judge), as well as a small-scale manual evaluation, to assess translation quality. We discuss the challenges of automatic MT evaluation, especially in the context of Arabic dialects. We also evaluate the ability of LLMs to classify the dialect used in a text. The study offers insights into the capabilities and limitations of current LLMs for dialectal Arabic machine translation, particularly highlighting the difficulty of handling dialectal diversity, although the results may be influenced by possible training data contamination, which is always a concern with LLMs.

This paper describes the CUNI submissions to the WMT25 General Translation task, namely for the English to Czech, English to Serbian, Czech to German and Czech to Ukrainian language pairs. We worked in multiple teams, each with a different approach, spanning from traditional, smaller Transformer NMT models trained on both sentence and document level, to fine-tuning LLMs using LoRA and CPO. We show that these methods are effective in improving automatic MT evaluation scores compared to the base pretrained models.

We present our IWSLT 2025 submission for the low-resource track on North Levantine Arabic to English speech translation, building on our IWSLT 2024 efforts. We retain last year’s cascade ASR architecture that combines a TDNN-F model and a Zipformer for the ASR step. We upgrade the Zipformer to the Zipformer-Large variant (253 M parameters vs. 66 M) to capture richer acoustic representations. For the MT part, to further alleviate data sparsity, we created a crowd-sourced parallel corpus covering five major Arabic dialects (Tunisian, Levantine, Moroccan, Algerian, Egyptian) curated via rigorous qualification and filtering. We show that using crowd-sourced data is feasible in low-resource scenarios as we observe improved automatic evaluation metrics across all dialects. We also experimented with the dataset under a high-resource scenario, where we had access to a large, high-quality Levantine Arabic corpus from LDC. In this setting, adding the crowd-sourced data does not improve the scores on the official validation set anymore. Our final submission scores 20.0 BLEU on the official test set.

In this paper, we present our submission for the token prediction task of EvaCun 2025. Our sys-tems are based on LLMs (Command-R, Mistral, and Aya Expanse) fine-tuned on the task data provided by the organizers. As we only pos-sess a very superficial knowledge of the subject field and the languages of the task, we simply used the training data without any task-specific adjustments, preprocessing, or filtering. We compare 3 different approaches (based on 3 different prompts) of obtaining the predictions, and we evaluate them on a held-out part of the data.

This paper presents the contributions of Charles University teams to the WMT24 General Translation task (English to Czech, German and Russian, and Czech to Ukrainian), and the WMT24 Translation into Low-Resource Languages of Spain task.Our most elaborate submission, CUNI-MH for en2cs, is the result of fine-tuning Mistral 7B v0.1 for translation using a three-stage process: Supervised fine-tuning using QLoRA, Contrastive Preference Optimization, and merging of model checkpoints. We also describe the CUNI-GA, CUNI-Transformer and CUNI-DocTransformer submissions, which are based on our systems from the previous year.Our en2ru system CUNI-DS uses a similar first stage as CUNI-MH (QLoRA for en2cs) and follows with transferring to en2ru.For en2de (CUNI-NL), we experimented with a LLM-based speech translation system, to translate without the speech input.For the Translation into Low-Resource Languages of Spain task, we performed QLoRA fine-tuning of a large LLM on a small amount of synthetic (backtranslated) data.

Building on a recent method for decoding translation candidates from a Machine Translation (MT) model via a genetic algorithm, we modify it to generate adversarial translations to test and challenge MT evaluation metrics. The produced translations score very well in an arbitrary MT evaluation metric selected beforehand, despite containing serious, deliberately introduced errors. The method can be used to create adversarial test sets to analyze the biases and shortcomings of the metrics. We publish various such test sets for the Czech to English language pair, as well as the code to convert any parallel data into a similar adversarial test set.

This paper presents ALADAN’s approach to the IWSLT 2024 Dialectal and Low-resource shared task, focusing on Levantine Arabic (apc) and Tunisian Arabic (aeb) to English speech translation (ST). Addressing challenges such as the lack of standardized orthography and limited training data, we propose a solution for data normalization in Dialectal Arabic, employing a modified Levenshtein distance and Word2vec models to find orthographic variants of the same word. Our system consists of a cascade ST system integrating two ASR systems (TDNN-F and Zipformer) and two NMT modules derived from pre-trained models (NLLB-200 1.3B distilled model and CohereAI’s Command-R). Additionally, we explore the integration of unsupervised textual and audio data, highlighting the importance of multi-dialectal datasets for both ASR and NMT tasks. Our system achieves BLEU score of 31.5 for Levantine Arabic on the official validation set.

This study examines neural machine translation (NMT) and its performance on texts that diverege from typical standards, focusing on how information is organized within sentences. We analyze surprisal distributions in source texts, human translations, and machine translations across several datasets to determine if NMT systems naturally promote a uniform density of surprisal in their translations, even when the original texts do not adhere to this principle.The findings reveal that NMT tends to align more closely with source texts in terms of surprisal uniformity compared to human translations.We analyzed absolute values of the surprisal uniformity measures as well, expecting that human translations will be less uniform. In contradiction to our initial hypothesis, we did not find comprehensive evidence for this claim, with some results suggesting this might be the case for very diverse texts, like poetry.

This paper presents the contributions of Charles University teams to the WMT23 General translation task (English to Czech and Czech to Ukrainian translation directions). Our main submission, CUNI-GA, is a result of applying a novel n-best list reranking and modification method on translation candidates produced by the two other submitted systems, CUNI-Transformer and CUNI-DocTransformer (document-level translation only used for the en → cs direction). Our method uses a genetic algorithm and MBR decoding to search for optimal translation under a given metric (in our case, a weighted combination of ChrF, BLEU, COMET22-DA, and COMET22-QE-DA). Our submissions are first in the constrained track and show competitive performance against top-tier unconstrained systems across various automatic metrics.

This paper explores negative lexical constraining in English to Czech neural machine translation. Negative lexical constraining is used to prohibit certain words or expressions in the translation produced by the NMT model. We compared various methods based on modifying either the decoding process or the training data. The comparison was performed on two tasks: paraphrasing and feedback-based translation refinement. We also studied how the methods “evade” the constraints, meaning that the disallowed expression is still present in the output, but in a changed form, most interestingly the case where a different surface form (for example different inflection) is produced. We propose a way to mitigate the issue through training with stemmed negative constraints, so that the ability of the model to induce different forms of a word might be used to prohibit the usage of all possible forms of the constraint. This helps to some extent, but the problem still persists in many cases.

We explore the effectiveness of character-level neural machine translation using Transformer architecture for various levels of language similarity and size of the training dataset. We evaluate the models using automatic MT metrics and show that translation between similar languages benefits from character-level input segmentation, while for less related languages, character-level vanilla Transformer-base often lags behind subword-level segmentation. We confirm previous findings that it is possible to close the gap by finetuning the already trained subword-level models to character-level.

We propose a genetic algorithm (GA) based method for modifying n-best lists produced by a machine translation (MT) system. Our method offers an innovative approach to improving MT quality and identifying weaknesses in evaluation metrics. Using common GA operations (mutation and crossover) on a list of hypotheses in combination with a fitness function (an arbitrary MT metric), we obtain novel and diverse outputs with high metric scores. With a combination of multiple MT metrics as the fitness function, the proposed method leads to an increase in translation quality as measured by other held-out automatic metrics.With a single metric (including popular ones such as COMET) as the fitness function, we find blind spots and flaws in the metric. This allows for an automated search for adversarial examples in an arbitrary metric, without prior assumptions on the form of such example. As a demonstration of the method, we create datasets of adversarial examples and use them to show that reference-free COMET is substantially less robust than the reference-based version.

We present the CUNI-Bergamot submission for the WMT22 General translation task. We compete in English-Czech direction. Our submission further explores block backtranslation techniques. Compared to the previous work, we measure performance in terms of COMET score and named entities translation accuracy. We evaluate performance of MBR decoding compared to traditional mixed backtranslation training and we show a possible synergy when using both of the techniques simultaneously. The results show that both approaches are effective means of improving translation quality and they yield even better results when combined.

This paper describes Charles University sub-mission for Terminology translation Shared Task at WMT21. The objective of this task is to design a system which translates certain terms based on a provided terminology database, while preserving high overall translation quality. We competed in English-French language pair. Our approach is based on providing the desired translations alongside the input sentence and training the model to use these provided terms. We lemmatize the terms both during the training and inference, to allow the model to learn how to produce correct surface forms of the words, when they differ from the forms provided in the terminology database. Our submission ranked second in Exact Match metric which evaluates the ability of the model to produce desired terms in the translation.

This paper describes Charles University sub-mission for Terminology translation shared task at WMT21. The objective of this task is to design a system which translates certain terms based on a provided terminology database, while preserving high overall translation quality. We competed in English-French language pair. Our approach is based on providing the desired translations alongside the input sentence and training the model to use these provided terms. We lemmatize the terms both during the training and inference, to allow the model to learn how to produce correct surface forms of the words, when they differ from the forms provided in the terminology database.

This work demonstrates that using the objective with independence assumption for modelling the span probability P (a_s , a_e ) = P (a_s )P (a_e) of span starting at position a_s and ending at position a_e has adverse effects. Therefore we propose multiple approaches to modelling joint probability P (a_s , a_e) directly. Among those, we propose a compound objective, composed from the joint probability while still keeping the objective with independence assumption as an auxiliary objective. We find that the compound objective is consistently superior or equal to other assumptions in exact match. Additionally, we identified common errors caused by the assumption of independence and manually checked the counterpart predictions, demonstrating the impact of the compound objective on the real examples. Our findings are supported via experiments with three extractive QA models (BIDAF, BERT, ALBERT) over six datasets and our code, individual results and manual analysis are available online.

Lexically constrained machine translation allows the user to manipulate the output sentence by enforcing the presence or absence of certain words and phrases. Although current approaches can enforce terms to appear in the translation, they often struggle to make the constraint word form agree with the rest of the generated output. Our manual analysis shows that 46% of the errors in the output of a baseline constrained model for English to Czech translation are related to agreement. We investigate mechanisms to allow neural machine translation to infer the correct word inflection given lemmatized constraints. In particular, we focus on methods based on training the model with constraints provided as part of the input sequence. Our experiments on English-Czech language pair show that this approach improves translation of constrained terms in both automatic and manual evaluation by reducing errors in agreement. Our approach thus eliminates inflection errors, without introducing new errors or decreasing overall quality of the translation.

This paper describes BUT-FIT’s submission at SemEval-2020 Task 5: Modelling Causal Reasoning in Language: Detecting Counterfactuals. The challenge focused on detecting whether a given statement contains a counterfactual (Subtask 1) and extracting both antecedent and consequent parts of the counterfactual from the text (Subtask 2). We experimented with various state-of-the-art language representation models (LRMs). We found RoBERTa LRM to perform the best in both subtasks. We achieved the first place in both exact match and F1 for Subtask 2 and ranked second for Subtask 1.

We participated in all three subtasks. In subtasks A and B, our submissions are based on pretrained language representation models (namely ALBERT) and data augmentation. We experimented with solving the task for another language, Czech, by means of multilingual models and machine translated dataset, or translated model inputs. We show that with a strong machine translation system, our system can be used in another language with a small accuracy loss. In subtask C, our submission, which is based on pretrained sequence-to-sequence model (BART), ranked 1st in BLEU score ranking, however, we show that the correlation between BLEU and human evaluation, in which our submission ended up 4th, is low. We analyse the metrics used in the evaluation and we propose an additional score based on model from subtask B, which correlates well with our manual ranking, as well as reranking method based on the same principle. We performed an error and dataset analysis for all subtasks and we present our findings.

This paper describes our system that was designed for Humor evaluation within the SemEval-2020 Task 7. The system is based on convolutional neural network architecture. We investigate the system on the official dataset, and we provide more insight to model itself to see how the learned inner features look.