Rico Sennrich

Rico Sennrich, Jannis Vamvas, Alireza Mohammadshahi

Hallucinations and off-target translation remain unsolved problems in MT, especially for low-resource languages and massively multilingual models. In this paper, we introduce two related methods to mitigate these failure cases with a modified decoding objective, without either requiring retraining or external models. In source-contrastive decoding, we search for a translation that is probable given the correct input, but improbable given a random input segment. In language-contrastive decoding, we search for a translation that is probable, but improbable given the wrong language indicator token. Experiments on the massively multilingual models M2M-100 (418M) and SMaLL-100 show that these methods suppress hallucinations and off-target translations, reducing the number of translations with segment-level chrF2 below 10 by 67-83% on average, and the number of translations with oscillatory hallucinations by 75-92% on average, across 57 tested translation directions. In a proof of concept on out-of-English translation, we also show that we can suppress off-target translations with large language models. We release our source code at https://github.com/ZurichNLP/ContraDecode.

Andrianos Michail, Elias Schuhmacher, Juri Opitz et al.

Dense retrieval models exhibit positional bias: retrieval effectiveness degrades when relevant information appears later in a passage (Zeng et al., 2025). We ask whether this bias can be reduced at inference time, without retraining and without sacrificing overall retrieval effectiveness. To this end, we adapt inference-time attention calibration (Schuhmacher et al., 2026) to downstream retrieval and extend it with a strength coefficient lambda that interpolates between the original and fully calibrated attention distributions. Across three embedding models on SQuAD-PosQ and FineWeb-PosQ, we examine how basket size, calibrated layer set, and strength affect the trade-off between positional fairness and retrieval effectiveness, finding that partial calibration frequently outperforms full calibration. A single configuration (B=128, lambda=0.5, 50% layer depth) improves the harmonic mean of nDCG@10 across positional groups on FineWeb-PosQ for all three models without per-model tuning, and applies to both <s>-pooled and last-token-pooled architectures. This default configuration transfers without modification to PosIR, which spans 10 languages and 31 domains, reducing the Position Sensitivity Index in all 16 length-quartile x model x retrieval-setting combinations, while preserving or improving aggregate nDCG@10. We release our extended codebase at https://github.com/impresso/fair-sentence-transformers

Noëmi Aepli, Chantal Amrhein, Florian Schottmann et al.

For sensible progress in natural language processing, it is important that we are aware of the limitations of the evaluation metrics we use. In this work, we evaluate how robust metrics are to non-standardized dialects, i.e. spelling differences in language varieties that do not have a standard orthography. To investigate this, we collect a dataset of human translations and human judgments for automatic machine translations from English to two Swiss German dialects. We further create a challenge set for dialect variation and benchmark existing metrics' performances. Our results show that existing metrics cannot reliably evaluate Swiss German text generation outputs, especially on segment level. We propose initial design adaptations that increase robustness in the face of non-standardized dialects, although there remains much room for further improvement. The dataset, code, and models are available here: https://github.com/textshuttle/dialect_eval

Jannis Vamvas, Johannes Graën, Rico Sennrich

We present SwissBERT, a masked language model created specifically for processing Switzerland-related text. SwissBERT is a pre-trained model that we adapted to news articles written in the national languages of Switzerland -- German, French, Italian, and Romansh. We evaluate SwissBERT on natural language understanding tasks related to Switzerland and find that it tends to outperform previous models on these tasks, especially when processing contemporary news and/or Romansh Grischun. Since SwissBERT uses language adapters, it may be extended to Swiss German dialects in future work. The model and our open-source code are publicly released at https://github.com/ZurichNLP/swissbert.

Jannis Vamvas, Rico Sennrich

Being able to rank the similarity of short text segments is an interesting bonus feature of neural machine translation. Translation-based similarity measures include direct and pivot translation probability, as well as translation cross-likelihood, which has not been studied so far. We analyze these measures in the common framework of multilingual NMT, releasing the NMTScore library (available at https://github.com/ZurichNLP/nmtscore). Compared to baselines such as sentence embeddings, translation-based measures prove competitive in paraphrase identification and are more robust against adversarial or multilingual input, especially if proper normalization is applied. When used for reference-based evaluation of data-to-text generation in 2 tasks and 17 languages, translation-based measures show a relatively high correlation to human judgments.

Hanxu Hu, Zdeněk Šnajdr, Pinzhen Chen et al.

Prior work has shown that large language models (LLMs) can translate unseen or low-resource languages by undergoing continued training or even by encoding a grammar book in their context. However, both methods typically overfit specific languages, with limited zero-shot transfer at test time. To translate extremely low-resource languages at scale, we argue that LLMs must acquire the meta-skill of utilizing in-context linguistic knowledge rather than memorizing specific languages. In this paper, we propose a reinforcement learning (RL) approach to unseen language translation given rich linguistic context, using a surface-level translation metric (chrF) as the reward. Empirically, despite the lightweight reward, our RL-trained models effectively extract and apply relevant linguistic information from the provided context, leading to better translations on completely unseen languages than in-context learning or supervised fine-tuning. Our analyses suggest that outcome-based RL can extend beyond conventional reasoning tasks like math and coding to serve as a recipe for language learning from context.

Genta Indra Winata, Alham Fikri Aji, Samuel Cahyawijaya et al.

Natural language processing (NLP) has a significant impact on society via technologies such as machine translation and search engines. Despite its success, NLP technology is only widely available for high-resource languages such as English and Chinese, while it remains inaccessible to many languages due to the unavailability of data resources and benchmarks. In this work, we focus on developing resources for languages in Indonesia. Despite being the second most linguistically diverse country, most languages in Indonesia are categorized as endangered and some are even extinct. We develop the first-ever parallel resource for 10 low-resource languages in Indonesia. Our resource includes datasets, a multi-task benchmark, and lexicons, as well as a parallel Indonesian-English dataset. We provide extensive analyses and describe the challenges when creating such resources. We hope that our work can spark NLP research on Indonesian and other underrepresented languages.

Guojun Wu, Shay B. Cohen, Rico Sennrich

We introduce a dataset comprising commercial machine translations, gathered weekly over six years across 12 translation directions. Since human A/B testing is commonly used, we assume commercial systems improve over time, which enables us to evaluate machine translation (MT) metrics based on their preference for more recent translations. Our study not only confirms several prior findings, such as the advantage of neural metrics over non-neural ones, but also explores the debated issue of how MT quality affects metric reliability--an investigation that smaller datasets in previous research could not sufficiently explore. Overall, our research demonstrates the dataset's value as a testbed for metric evaluation. We release our code at https://github.com/gjwubyron/Evo

Joris Baan, Nico Daheim, Evgenia Ilia et al.

Recent advances of powerful Language Models have allowed Natural Language Generation (NLG) to emerge as an important technology that can not only perform traditional tasks like summarisation or translation, but also serve as a natural language interface to a variety of applications. As such, it is crucial that NLG systems are trustworthy and reliable, for example by indicating when they are likely to be wrong; and supporting multiple views, backgrounds and writing styles -- reflecting diverse human sub-populations. In this paper, we argue that a principled treatment of uncertainty can assist in creating systems and evaluation protocols better aligned with these goals. We first present the fundamental theory, frameworks and vocabulary required to represent uncertainty. We then characterise the main sources of uncertainty in NLG from a linguistic perspective, and propose a two-dimensional taxonomy that is more informative and faithful than the popular aleatoric/epistemic dichotomy. Finally, we move from theory to applications and highlight exciting research directions that exploit uncertainty to power decoding, controllable generation, self-assessment, selective answering, active learning and more.

Biao Zhang, Barry Haddow, Rico Sennrich

End-to-end (E2E) speech-to-text translation (ST) often depends on pretraining its encoder and/or decoder using source transcripts via speech recognition or text translation tasks, without which translation performance drops substantially. However, transcripts are not always available, and how significant such pretraining is for E2E ST has rarely been studied in the literature. In this paper, we revisit this question and explore the extent to which the quality of E2E ST trained on speech-translation pairs alone can be improved. We reexamine several techniques proven beneficial to ST previously, and offer a set of best practices that biases a Transformer-based E2E ST system toward training from scratch. Besides, we propose parameterized distance penalty to facilitate the modeling of locality in the self-attention model for speech. On four benchmarks covering 23 languages, our experiments show that, without using any transcripts or pretraining, the proposed system reaches and even outperforms previous studies adopting pretraining, although the gap remains in (extremely) low-resource settings. Finally, we discuss neural acoustic feature modeling, where a neural model is designed to extract acoustic features from raw speech signals directly, with the goal to simplify inductive biases and add freedom to the model in describing speech. For the first time, we demonstrate its feasibility and show encouraging results on ST tasks.

Jannis Vamvas, Rico Sennrich

Omission and addition of content is a typical issue in neural machine translation. We propose a method for detecting such phenomena with off-the-shelf translation models. Using contrastive conditioning, we compare the likelihood of a full sequence under a translation model to the likelihood of its parts, given the corresponding source or target sequence. This allows to pinpoint superfluous words in the translation and untranslated words in the source even in the absence of a reference translation. The accuracy of our method is comparable to a supervised method that requires a custom quality estimation model.

Zifan Jiang, Gerard Sant, Amit Moryossef et al.

We present SignCLIP, which re-purposes CLIP (Contrastive Language-Image Pretraining) to project spoken language text and sign language videos, two classes of natural languages of distinct modalities, into the same space. SignCLIP is an efficient method of learning useful visual representations for sign language processing from large-scale, multilingual video-text pairs, without directly optimizing for a specific task or sign language which is often of limited size. We pretrain SignCLIP on Spreadthesign, a prominent sign language dictionary consisting of ~500 thousand video clips in up to 44 sign languages, and evaluate it with various downstream datasets. SignCLIP discerns in-domain signing with notable text-to-video/video-to-text retrieval accuracy. It also performs competitively for out-of-domain downstream tasks such as isolated sign language recognition upon essential few-shot prompting or fine-tuning. We analyze the latent space formed by the spoken language text and sign language poses, which provides additional linguistic insights. Our code and models are openly available.

Farhad Nooralahzadeh, Rico Sennrich

While several benefits were realized for multilingual vision-language pretrained models, recent benchmarks across various tasks and languages showed poor cross-lingual generalisation when multilingually pre-trained vision-language models are applied to non-English data, with a large gap between (supervised) English performance and (zero-shot) cross-lingual transfer. In this work, we explore the poor performance of these models on a zero-shot cross-lingual visual question answering (VQA) task, where models are fine-tuned on English visual-question data and evaluated on 7 typologically diverse languages. We improve cross-lingual transfer with three strategies: (1) we introduce a linguistic prior objective to augment the cross-entropy loss with a similarity-based loss to guide the model during training, (2) we learn a task-specific subnetwork that improves cross-lingual generalisation and reduces variance without model modification, (3) we augment training examples using synthetic code-mixing to promote alignment of embeddings between source and target languages. Our experiments on xGQA using the pretrained multilingual multimodal transformers UC2 and M3P demonstrate the consistent effectiveness of the proposed fine-tuning strategy for 7 languages, outperforming existing transfer methods with sparse models. Code and data to reproduce our findings are publicly available.

Jason Armitage, Leonardo Impett, Rico Sennrich

In a busy city street, a pedestrian surrounded by distractions can pick out a single sign if it is relevant to their route. Artificial agents in outdoor Vision-and-Language Navigation (VLN) are also confronted with detecting supervisory signal on environment features and location in inputs. To boost the prominence of relevant features in transformer-based architectures without costly preprocessing and pretraining, we take inspiration from priority maps - a mechanism described in neuropsychological studies. We implement a novel priority map module and pretrain on auxiliary tasks using low-sample datasets with high-level representations of routes and environment-related references to urban features. A hierarchical process of trajectory planning - with subsequent parameterised visual boost filtering on visual inputs and prediction of corresponding textual spans - addresses the core challenges of cross-modal alignment and feature-level localisation. The priority map module is integrated into a feature-location framework that doubles the task completion rates of standalone transformers and attains state-of-the-art performance on the Touchdown benchmark for VLN. Code and data are referenced in Appendix C.

Biao Zhang, Barry Haddow, Rico Sennrich

For end-to-end speech translation, regularizing the encoder with the Connectionist Temporal Classification (CTC) objective using the source transcript or target translation as labels can greatly improve quality metrics. However, CTC demands an extra prediction layer over the vocabulary space, bringing in nonnegligible model parameters and computational overheads, although this layer is typically not used for inference. In this paper, we re-examine the need for genuine vocabulary labels for CTC for regularization and explore strategies to reduce the CTC label space, targeting improved efficiency without quality degradation. We propose coarse labeling for CTC (CoLaCTC), which merges vocabulary labels via simple heuristic rules, such as using truncation, division or modulo (MOD) operations. Despite its simplicity, our experiments on 4 source and 8 target languages show that CoLaCTC with MOD particularly can compress the label space aggressively to 256 and even further, gaining training efficiency (1.18x ~ 1.77x speedup depending on the original vocabulary size) yet still delivering comparable or better performance than the CTC baseline. We also show that CoLaCTC successfully generalizes to CTC regularization regardless of using transcript or translation for labeling.

Michelle Wastl, Jannis Vamvas, Rico Sennrich

Recognizing semantic differences across documents, especially in different languages, is crucial for text generation evaluation and multilingual content alignment. However, as a standalone task it has received little attention. We address this by introducing SwissGov-RSD, the first naturalistic, document-level, cross-lingual dataset for semantic difference recognition. It encompasses a total of 224 multi-parallel documents in English-German, English-French, and English-Italian with token-level difference annotations by human annotators. We evaluate a variety of open-source and closed source large language models as well as encoder models across different fine-tuning settings on this new benchmark. Our results show that current automatic approaches perform poorly compared to their performance on monolingual, sentence-level, and synthetic benchmarks, revealing a considerable gap for both LLMs and encoder models. We make our code and datasets publicly available.

Elias Schuhmacher, Andrianos Michail, Juri Opitz et al.

To be discoverable in an embedding-based search process, each part of a document should be reflected in its embedding representation. To quantify any potential reflection biases, we introduce a permutation-based evaluation framework. With this, we observe that state-of-the-art embedding models exhibit systematic positional and language biases when documents are longer and consist of multiple segments. Specifically, early segments and segments in higher-resource languages like English are over-represented, while later segments and segments in lower-resource languages are marginalized. In our further analysis, we find that the positional bias stems from front-loaded attention distributions in pooling-token embeddings, where early tokens receive more attention. To mitigate this issue, we introduce an inference-time attention calibration method that redistributes attention more evenly across document positions, increasing discoverabiltiy of later segments. Our evaluation framework and attention calibration is available at https://github.com/impresso/fair-sentence-transformers

Hanxu Hu, Yuxuan Wang, Maggie Huan et al.

Reinforcement learning with Verifiable Rewards (RLVR) has emerged as a powerful paradigm for eliciting reasoning capabilities in large language models, particularly in mathematics and coding. While recent efforts have extended this paradigm to broader general scientific (STEM) domains, the complex interplay between supervised fine-tuning (SFT) and RL in these contexts remains underexplored. In this paper, we conduct controlled experiments revealing a critical challenge: for general STEM domains, RL applied directly to base models is highly sample-inefficient and is consistently surpassed by supervised fine-tuning (SFT) on moderate-quality responses. Yet sequential SFT followed by RL can further improve performance, suggesting that the two stages play complementary roles, and that how training data is allocated between them matters. Therefore, we propose DeReason, a difficulty-based data decoupling strategy for general reasoning. DeReason partitions training data by reasoning intensity estimated via LLM-based scoring into reasoning-intensive and non-reasoning-intensive subsets. It allocates broad-coverage, non-reasoning-intensive problems to SFT to establish foundational domain knowledge, and reserves a focused subset of difficult problems for RL to cultivate complex reasoning. We demonstrate that this principled decoupling yields better performance than randomly splitting the data for sequential SFT and RL. Extensive experiments on general STEM and mathematical benchmarks demonstrate that our decoupled curriculum training significantly outperforms SFT-only, RL-only, and random-split baselines. Our work provides a systematic study of the interplay between SFT and RL for general reasoning, offering a highly effective and generalized post-training recipe.

Tannon Kew, Florian Schottmann, Rico Sennrich

The vast majority of today's large language models (LLMs) are English-centric, having been pretrained predominantly on English text. Yet, in order to meet user expectations, models need to be able to respond appropriately in multiple languages once deployed in downstream applications. This requires strong cross-lingual transfer abilities. In this work, we investigate the minimal amount of multilinguality required during finetuning to elicit cross-lingual generalisation in English-centric LLMs. In experiments across four LLMs, we find that multilingual instruction tuning with as few as two to three languages is both necessary and sufficient to elicit effective cross-lingual generalisation, with the limiting factor being the degree to which a target language is seen during pretraining. Evaluations on five different tasks further reveal that multilingual instruction tuning is most beneficial for generative tasks that assume input/output language agreement, such as in chat settings, while being of less importance for highly structured classification-style tasks. Our code and data is available at https://github.com/ZurichNLP/multilingual-instruction-tuning.

Jannis Vamvas, Rico Sennrich

Minimum Bayes Risk (MBR) decoding is a text generation technique that has been shown to improve the quality of machine translations, but is expensive, even if a sampling-based approximation is used. Besides requiring a large number of sampled sequences, it requires the pairwise calculation of a utility metric, which has quadratic complexity. In this paper, we propose to approximate pairwise metric scores with scores calculated against aggregated reference representations. This changes the complexity of utility estimation from $O(n^2)$ to $O(n)$, while empirically preserving most of the quality gains of MBR decoding. We release our source code at https://github.com/ZurichNLP/mbr

Alireza Mohammadshahi, Jannis Vamvas, Rico Sennrich

Massively multilingual machine translation models allow for the translation of a large number of languages with a single model, but have limited performance on low- and very-low-resource translation directions. Pivoting via high-resource languages remains a strong strategy for low-resource directions, and in this paper we revisit ways of pivoting through multiple languages. Previous work has used a simple averaging of probability distributions from multiple paths, but we find that this performs worse than using a single pivot, and exacerbates the hallucination problem because the same hallucinations can be probable across different paths. We also propose MaxEns, a novel combination strategy that makes the output biased towards the most confident predictions, hypothesising that confident predictions are less prone to be hallucinations. We evaluate different strategies on the FLORES benchmark for 20 low-resource language directions, demonstrating that MaxEns improves translation quality for low-resource languages while reducing hallucination in translations, compared to both direct translation and an averaging approach. On average, multi-pivot strategies still lag behind using English as a single pivot language, raising the question of how to identify the best pivoting strategy for a given translation direction.

Andrianos Michail, Simon Clematide, Rico Sennrich

The evaluation of cross-lingual semantic search models is often limited to existing datasets from tasks such as information retrieval and semantic textual similarity. We introduce Cross-Lingual Semantic Discrimination (CLSD), a lightweight evaluation task that requires only parallel sentences and a Large Language Model (LLM) to generate adversarial distractors. CLSD measures an embedding model's ability to rank the true parallel sentence above semantically misleading but lexically similar alternatives. As a case study, we construct CLSD datasets for German--French in the news domain. Our experiments show that models fine-tuned for retrieval tasks benefit from pivoting through English, whereas bitext mining models perform best in direct cross-lingual settings. A fine-grained similarity analysis further reveals that embedding models differ in their sensitivity to linguistic perturbations. We release our code and datasets under AGPL-3.0: https://github.com/impresso/cross_lingual_semantic_discrimination

Aso Mahmudi, Sina Ahmadi, Kemal Kurniawan et al.

Collecting and annotating morphological data present significant challenges, requiring linguistic expertise, methodological rigour, and substantial resources. These barriers are particularly acute for low-resource languages and varieties. To accelerate this process, we introduce \texttt{CommonMorph}, a comprehensive platform that streamlines morphological data collection development through a three-tiered approach: expert linguistic definition, contributor elicitation, and community validation. The platform minimises manual work by incorporating active learning, annotation suggestions, and tools to import and adapt materials from related languages. It accommodates diverse morphological systems, including fusional, agglutinative, and root-and-pattern morphologies. Its open-source design and UniMorph-compatible outputs ensure accessibility and interoperability with NLP tools. Our platform is accessible at https://common-morph.com, offering a replicable model for preserving linguistic diversity through collaborative technology.

Zifan Jiang, Colin Leong, Amit Moryossef et al.

We present a comprehensive study on meaningfully evaluating sign language utterances in the form of human skeletal poses. The study covers keypoint distance-based, embedding-based, and back-translation-based metrics. We show tradeoffs between different metrics in different scenarios through automatic meta-evaluation of sign-level retrieval and a human correlation study of text-to-pose translation across different sign languages. Our findings and the open-source pose-evaluation toolkit provide a practical and reproducible way of developing and evaluating sign language translation or generation systems.

Jannis Vamvas, Ignacio Pérez Prat, Angela Heldstab et al.

Recent strategies for low-resource machine translation rely on LLMs to generate synthetic data from higher-resource languages. We find that this method fails for Romansh, because LLMs tend to confuse its 6 distinct language varieties. Our experiments show that instead, the direction of data augmentation should be aligned with the resource gradient between source and target language. This approach surpasses Gemini 3 Pro in the lowest-resource variety of Romansh by 23 BLEU. A human evaluation confirms that our experiments yield the first model that generates fluent translations in the individual Romansh varieties.

Kirill Semenov, Rico Sennrich

For multilingual factual knowledge assessment of LLMs, benchmarks such as MLAMA use template translations that do not take into account the grammatical and semantic information of the named entities inserted in the sentence. This leads to numerous instances of ungrammaticality or wrong wording of the final prompts, which complicates the interpretation of scores, especially for languages that have a rich morphological inventory. In this work, we sample 4 Slavic languages from the MLAMA dataset and compare the knowledge retrieval scores between the initial (templated) MLAMA dataset and its sentence-level translations made by Google Translate and ChatGPT. We observe a significant increase in knowledge retrieval scores, and provide a qualitative analysis for possible reasons behind it. We also make an additional analysis of 5 more languages from different families and see similar patterns. Therefore, we encourage the community to control the grammaticality of highly multilingual datasets for higher and more interpretable results, which is well approximated by whole sentence translation with neural MT or LLM systems. The dataset and all related code is published at the Github repository: https://github.com/ZurichNLP/Fluent-mLAMA.

Jannis Vamvas, Noëmi Aepli, Rico Sennrich

Creating neural text encoders for written Swiss German is challenging due to a dearth of training data combined with dialectal variation. In this paper, we build on several existing multilingual encoders and adapt them to Swiss German using continued pre-training. Evaluation on three diverse downstream tasks shows that simply adding a Swiss German adapter to a modular encoder achieves 97.5% of fully monolithic adaptation performance. We further find that for the task of retrieving Swiss German sentences given Standard German queries, adapting a character-level model is more effective than the other adaptation strategies. We release our code and the models trained for our experiments at https://github.com/ZurichNLP/swiss-german-text-encoders

Michelle Wastl, Jannis Vamvas, Rico Sennrich

Detecting the translation direction of parallel text has applications for machine translation training and evaluation, but also has forensic applications such as resolving plagiarism or forgery allegations. In this work, we explore an unsupervised approach to translation direction detection based on the simple hypothesis that $p(\text{translation}|\text{original})>p(\text{original}|\text{translation})$, motivated by the well-known simplification effect in translationese or machine-translationese. In experiments with massively multilingual machine translation models across 20 translation directions, we confirm the effectiveness of the approach for high-resource language pairs, achieving document-level accuracies of 82--96% for NMT-produced translations, and 60--81% for human translations, depending on the model used. Code and demo are available at https://github.com/ZurichNLP/translation-direction-detection

Jannis Vamvas, Rico Sennrich

Automatically highlighting words that cause semantic differences between two documents could be useful for a wide range of applications. We formulate recognizing semantic differences (RSD) as a token-level regression task and study three unsupervised approaches that rely on a masked language model. To assess the approaches, we begin with basic English sentences and gradually move to more complex, cross-lingual document pairs. Our results show that an approach based on word alignment and sentence-level contrastive learning has a robust correlation to gold labels. However, all unsupervised approaches still leave a large margin of improvement. Code to reproduce our experiments is available at https://github.com/ZurichNLP/recognizing-semantic-differences

Biao Zhang, Mathias Müller, Rico Sennrich

Despite recent successes with neural models for sign language translation (SLT), translation quality still lags behind spoken languages because of the data scarcity and modality gap between sign video and text. To address both problems, we investigate strategies for cross-modality representation sharing for SLT. We propose SLTUNET, a simple unified neural model designed to support multiple SLTrelated tasks jointly, such as sign-to-gloss, gloss-to-text and sign-to-text translation. Jointly modeling different tasks endows SLTUNET with the capability to explore the cross-task relatedness that could help narrow the modality gap. In addition, this allows us to leverage the knowledge from external resources, such as abundant parallel data used for spoken-language machine translation (MT). We show in experiments that SLTUNET achieves competitive and even state-of-the-art performance on PHOENIX-2014T and CSL-Daily when augmented with MT data and equipped with a set of optimization techniques. We further use the DGS Corpus for end-to-end SLT for the first time. It covers broader domains with a significantly larger vocabulary, which is more challenging and which we consider to allow for a more realistic assessment of the current state of SLT than the former two. Still, SLTUNET obtains improved results on the DGS Corpus. Code is available at https://github.com/bzhangGo/sltunet.

Biao Zhang, Ivan Titov, Rico Sennrich

Independence assumptions during sequence generation can speed up inference, but parallel generation of highly inter-dependent tokens comes at a cost in quality. Instead of assuming independence between neighbouring tokens (semi-autoregressive decoding, SA), we take inspiration from bidirectional sequence generation and introduce a decoder that generates target words from the left-to-right and right-to-left directions simultaneously. We show that we can easily convert a standard architecture for unidirectional decoding into a bidirectional decoder by simply interleaving the two directions and adapting the word positions and self-attention masks. Our interleaved bidirectional decoder (IBDecoder) retains the model simplicity and training efficiency of the standard Transformer, and on five machine translation tasks and two document summarization tasks, achieves a decoding speedup of ~2X compared to autoregressive decoding with comparable quality. Notably, it outperforms left-to-right SA because the independence assumptions in IBDecoder are more felicitous. To achieve even higher speedups, we explore hybrid models where we either simultaneously predict multiple neighbouring tokens per direction, or perform multi-directional decoding by partitioning the target sequence. These methods achieve speedups to 4X-11X across different tasks at the cost of <1 BLEU or <0.5 ROUGE (on average). Source code is released at https://github.com/bzhangGo/zero.

Biao Zhang, Rico Sennrich

Layer normalization (LayerNorm) has been successfully applied to various deep neural networks to help stabilize training and boost model convergence because of its capability in handling re-centering and re-scaling of both inputs and weight matrix. However, the computational overhead introduced by LayerNorm makes these improvements expensive and significantly slows the underlying network, e.g. RNN in particular. In this paper, we hypothesize that re-centering invariance in LayerNorm is dispensable and propose root mean square layer normalization, or RMSNorm. RMSNorm regularizes the summed inputs to a neuron in one layer according to root mean square (RMS), giving the model re-scaling invariance property and implicit learning rate adaptation ability. RMSNorm is computationally simpler and thus more efficient than LayerNorm. We also present partial RMSNorm, or pRMSNorm where the RMS is estimated from p% of the summed inputs without breaking the above properties. Extensive experiments on several tasks using diverse network architectures show that RMSNorm achieves comparable performance against LayerNorm but reduces the running time by 7%~64% on different models. Source code is available at https://github.com/bzhangGo/rmsnorm.

Antonio Valerio Miceli Barone, Rico Sennrich

Automated documentation of programming source code and automated code generation from natural language are challenging tasks of both practical and scientific interest. Progress in these areas has been limited by the low availability of parallel corpora of code and natural language descriptions, which tend to be small and constrained to specific domains. In this work we introduce a large and diverse parallel corpus of a hundred thousands Python functions with their documentation strings ("docstrings") generated by scraping open source repositories on GitHub. We describe baseline results for the code documentation and code generation tasks obtained by neural machine translation. We also experiment with data augmentation techniques to further increase the amount of training data. We release our datasets and processing scripts in order to stimulate research in these areas.

Rico Sennrich, Barry Haddow

Neural machine translation has recently achieved impressive results, while using little in the way of external linguistic information. In this paper we show that the strong learning capability of neural MT models does not make linguistic features redundant; they can be easily incorporated to provide further improvements in performance. We generalize the embedding layer of the encoder in the attentional encoder--decoder architecture to support the inclusion of arbitrary features, in addition to the baseline word feature. We add morphological features, part-of-speech tags, and syntactic dependency labels as input features to English<->German, and English->Romanian neural machine translation systems. In experiments on WMT16 training and test sets, we find that linguistic input features improve model quality according to three metrics: perplexity, BLEU and CHRF3. An open-source implementation of our neural MT system is available, as are sample files and configurations.

Nikita Moghe, Arnisa Fazla, Chantal Amrhein et al. · microsoft-research

Recent machine translation (MT) metrics calibrate their effectiveness by correlating with human judgement but without any insights about their behaviour across different error types. Challenge sets are used to probe specific dimensions of metric behaviour but there are very few such datasets and they either focus on a limited number of phenomena or a limited number of language pairs. We introduce ACES, a contrastive challenge set spanning 146 language pairs, aimed at discovering whether metrics can identify 68 translation accuracy errors. These phenomena range from simple alterations at the word/character level to more complex errors based on discourse and real-world knowledge. We conduct a large-scale study by benchmarking ACES on 50 metrics submitted to the WMT 2022 and 2023 metrics shared tasks. We benchmark metric performance, assess their incremental performance over successive campaigns, and measure their sensitivity to a range of linguistic phenomena. We also investigate claims that Large Language Models (LLMs) are effective as MT evaluators by evaluating on ACES. Our results demonstrate that different metric families struggle with different phenomena and that LLM-based methods fail to demonstrate reliable performance. Our analyses indicate that most metrics ignore the source sentence, tend to prefer surface-level overlap and end up incorporating properties of base models which are not always beneficial. We expand ACES to include error span annotations, denoted as SPAN-ACES and we use this dataset to evaluate span-based error metrics showing these metrics also need considerable improvement. Finally, we provide a set of recommendations for building better MT metrics, including focusing on error labels instead of scores, ensembling, designing strategies to explicitly focus on the source sentence, focusing on semantic content and choosing the right base model for representations.

Xiao Zhu, Chenmien Tan, Pinzhen Chen et al.

Reward models (RMs) play a crucial role in Reinforcement Learning from Human Feedback by serving as proxies for human preferences in aligning large language models. In this paper, we identify a model preference bias in RMs, where they systematically assign disproportionately high scores to responses from certain policy models. This bias distorts ranking evaluations and leads to unfair judgments. To address this issue, we propose a calibration method named CHatbot Arena calibrated Reward Modeling (CHARM) that leverages Elo scores from the Chatbot Arena leaderboard to mitigate RM overvaluation. We also introduce a Mismatch Degree metric to measure this preference bias. Our approach is computationally efficient, requiring only a small preference dataset for continued training of the RM. We conduct extensive experiments on reward model benchmarks and human preference alignment. Results demonstrate that our calibrated RMs (1) achieve improved evaluation accuracy on RM-Bench and the Chat-Hard domain of RewardBench, and (2) exhibit a stronger correlation with human preferences by producing scores more closely aligned with Elo rankings. By mitigating model preference bias, our method provides a generalizable and efficient solution for building fairer and more reliable reward models.

Negar Foroutan, Clara Meister, Debjit Paul et al.

Tokenization is the first -- and often least scrutinized -- step of most NLP pipelines. Standard algorithms for learning tokenizers rely on frequency-based objectives, which favor languages dominant in the training data and consequently leave lower-resource languages with tokenizations that are disproportionately longer, morphologically implausible, or even riddled with <UNK> placeholders. This phenomenon ultimately amplifies computational and financial inequalities between users from different language backgrounds. To remedy this, we introduce Parity-aware Byte Pair Encoding (BPE), a variant of the widely-used BPE algorithm. At every merge step, Parity-aware BPE maximizes the compression gain of the currently worst-compressed language, trading a small amount of global compression for cross-lingual parity. We find empirically that Parity-aware BPE leads to more equitable token counts across languages, with negligible impact on global compression rate and no substantial effect on language-model performance in downstream tasks.

Michelle Wastl, Jannis Vamvas, Selena Calleri et al.

We present 20min-XD (20 Minuten cross-lingual document-level), a French-German, document-level comparable corpus of news articles, sourced from the Swiss online news outlet 20 Minuten/20 minutes. Our dataset comprises around 15,000 article pairs spanning 2015 to 2024, automatically aligned based on semantic similarity. We detail the data collection process and alignment methodology. Furthermore, we provide a qualitative and quantitative analysis of the corpus. The resulting dataset exhibits a broad spectrum of cross-lingual similarity, ranging from near-translations to loosely related articles, making it valuable for various NLP applications and broad linguistically motivated studies. We publicly release the dataset in document- and sentence-aligned versions and code for the described experiments.

Marco Cognetta, Tatsuya Hiraoka, Naoaki Okazaki et al.

We explore threshold vocabulary trimming in Byte-Pair Encoding subword tokenization, a postprocessing step that replaces rare subwords with their component subwords. The technique is available in popular tokenization libraries but has not been subjected to rigorous scientific scrutiny. While the removal of rare subwords is suggested as best practice in machine translation implementations, both as a means to reduce model size and for improving model performance through robustness, our experiments indicate that, across a large space of hyperparameter settings, vocabulary trimming fails to improve performance, and is even prone to incurring heavy degradation.

Hanxu Hu, Jannis Vamvas, Rico Sennrich

LLMs have paved the way for truly simple document-level machine translation, but challenges such as omission errors remain. In this paper, we study a simple method for handling document-level machine translation, by leveraging previous contexts in a multi-turn conversational manner. Specifically, by decomposing documents into segments and iteratively translating them while maintaining previous turns, this method ensures coherent translations without additional training, and can fully re-use the KV cache of previous turns thus minimizing computational overhead. We further propose a `source-primed' method that first provides the whole source document before multi-turn translation. We empirically show this multi-turn method outperforms both translating entire documents in a single turn and translating each segment independently according to multiple automatic metrics in representative LLMs, establishing a strong baseline for document-level translation using LLMs.

Jannis Vamvas, Ignacio Pérez Prat, Not Battesta Soliva et al.

The Romansh language, spoken in Switzerland, has limited resources for machine translation evaluation. In this paper, we present a benchmark for six varieties of Romansh: Rumantsch Grischun, a supra-regional variety, and five regional varieties: Sursilvan, Sutsilvan, Surmiran, Puter, and Vallader. Our reference translations were created by human translators based on the WMT24++ benchmark, which ensures parallelism with more than 55 other languages. An automatic evaluation of existing MT systems and LLMs shows that translation out of Romansh into German is handled relatively well for all the varieties, but translation into Romansh is still challenging.

Hanxu Hu, Xingxing Zhang, Jannis Vamvas et al.

Large Language Models have achieved strong performance on reasoning tasks, solving competition-level coding and math problems. However, their scalability is limited by human-labeled datasets and the lack of large-scale, challenging coding problem training data. Existing competitive coding datasets contain only thousands to tens of thousands of problems. Previous synthetic data generation methods rely on either augmenting existing instruction datasets or selecting challenging problems from human-labeled data. In this paper, we propose QueST, a novel framework which combines difficulty-aware graph sampling and difficulty-aware rejection fine-tuning that directly optimizes specialized generators to create challenging coding problems. Our trained generators demonstrate superior capability compared to even GPT-4o at creating challenging problems that benefit downstream performance. We leverage QueST to generate large-scale synthetic coding problems, which we then use to distill from strong teacher models with long chain-of-thought or to conduct reinforcement learning for smaller models, proving effective in both scenarios. Our distillation experiments demonstrate significant performance gains. Specifically, after fine-tuning Qwen3-8B-base on 100K difficult problems generated by QueST, we surpass the performance of the original Qwen3-8B on LiveCodeBench. With an additional 112K examples (i.e., 28K human-written problems paired with multiple synthetic solutions), our 8B model matches the performance of the much larger DeepSeek-R1-671B. These findings indicate that generating complex problems via QueST offers an effective and scalable approach to advancing the frontiers of competitive coding and reasoning for large language models.

Ahmet Yavuz Uluslu, Tannon Kew, Tilia Ellendorff et al.

Large language models (LLMs) often achieve high performance in native language identification (NLI) benchmarks by leveraging superficial contextual clues such as names, locations, and cultural stereotypes, rather than the underlying linguistic patterns indicative of native language (L1) influence. To improve robustness, previous work has instructed LLMs to disregard such clues. In this work, we demonstrate that such a strategy is unreliable and model predictions can be easily altered by misleading hints. To address this problem, we introduce an agentic NLI pipeline inspired by forensic linguistics, where specialized agents accumulate and categorize diverse linguistic evidence before an independent final overall assessment. In this final assessment, a goal-aware coordinating agent synthesizes all evidence to make the NLI prediction. On two benchmark datasets, our approach significantly enhances NLI robustness against misleading contextual clues and performance consistency compared to standard prompting methods.

Gerard Sant, Zifan Jiang, Carlos Escolano et al.

In recent years, sign language processing (SLP) has gained importance in the general field of Natural Language Processing. However, compared to research on spoken languages, SLP research is hindered by complex ad-hoc code, inadvertently leading to low reproducibility and unfair comparisons. Existing tools that are built for fast and reproducible experimentation, such as Hugging Face, are not flexible enough to seamlessly integrate sign language experiments. This view is confirmed by a survey we conducted among SLP researchers. To address these challenges, we introduce MultimodalHugs, a framework built on top of Hugging Face that enables more diverse data modalities and tasks, while inheriting the well-known advantages of the Hugging Face ecosystem. Even though sign languages are our primary focus, MultimodalHugs adds a layer of abstraction that makes it more widely applicable to other use cases that do not fit one of the standard templates of Hugging Face. We provide quantitative experiments to illustrate how MultimodalHugs can accommodate diverse modalities such as pose estimation data for sign languages, or pixel data for text characters.

Zachary Hopton, Jannis Vamvas, Andrin Büchler et al.

The five idioms (i.e., varieties) of the Romansh language are largely standardized and are taught in the schools of the respective communities in Switzerland. In this paper, we present the first parallel corpus of Romansh idioms. The corpus is based on 291 schoolbook volumes, which are comparable in content for the five idioms. We use automatic alignment methods to extract 207k multi-parallel segments from the books, with more than 2M tokens in total. A small-scale human evaluation confirms that the segments are highly parallel, making the dataset suitable for NLP applications such as machine translation between Romansh idioms. We release the parallel and unaligned versions of the dataset under a CC-BY-NC-SA license and demonstrate its utility for machine translation by training and evaluating an LLM on a sample of the dataset.

Patrick Haller, Jannis Vamvas, Rico Sennrich et al.

A growing body of work has been querying LLMs with political questions to evaluate their potential biases. However, this probing method has limited stability, making comparisons between models unreliable. In this paper, we argue that LLMs need more context. We propose a new probing task, Questionnaire Modeling (QM), that uses human survey data as in-context examples. We show that QM improves the stability of question-based bias evaluation, and demonstrate that it may be used to compare instruction-tuned models to their base versions. Experiments with LLMs of various sizes indicate that instruction tuning can indeed change the direction of bias. Furthermore, we observe a trend that larger models are able to leverage in-context examples more effectively, and generally exhibit smaller bias scores in QM. Data and code are publicly available.

Chantal Amrhein, Florian Schottmann, Rico Sennrich et al.

Natural language generation models reproduce and often amplify the biases present in their training data. Previous research explored using sequence-to-sequence rewriting models to transform biased model outputs (or original texts) into more gender-fair language by creating pseudo training data through linguistic rules. However, this approach is not practical for languages with more complex morphology than English. We hypothesise that creating training data in the reverse direction, i.e. starting from gender-fair text, is easier for morphologically complex languages and show that it matches the performance of state-of-the-art rewriting models for English. To eliminate the rule-based nature of data creation, we instead propose using machine translation models to create gender-biased text from real gender-fair text via round-trip translation. Our approach allows us to train a rewriting model for German without the need for elaborate handcrafted rules. The outputs of this model increased gender-fairness as shown in a human evaluation study.

Simone Tedeschi, Johan Bos, Thierry Declerck et al.

In the last five years, there has been a significant focus in Natural Language Processing (NLP) on developing larger Pretrained Language Models (PLMs) and introducing benchmarks such as SuperGLUE and SQuAD to measure their abilities in language understanding, reasoning, and reading comprehension. These PLMs have achieved impressive results on these benchmarks, even surpassing human performance in some cases. This has led to claims of superhuman capabilities and the provocative idea that certain tasks have been solved. In this position paper, we take a critical look at these claims and ask whether PLMs truly have superhuman abilities and what the current benchmarks are really evaluating. We show that these benchmarks have serious limitations affecting the comparison between humans and PLMs and provide recommendations for fairer and more transparent benchmarks.

Chantal Amrhein, Rico Sennrich

Neural metrics have achieved impressive correlation with human judgements in the evaluation of machine translation systems, but before we can safely optimise towards such metrics, we should be aware of (and ideally eliminate) biases toward bad translations that receive high scores. Our experiments show that sample-based Minimum Bayes Risk decoding can be used to explore and quantify such weaknesses. When applying this strategy to COMET for en-de and de-en, we find that COMET models are not sensitive enough to discrepancies in numbers and named entities. We further show that these biases are hard to fully remove by simply training on additional synthetic data and release our code and data for facilitating further experiments.

Antonio Valerio Miceli-Barone, Alexandra Birch, Rico Sennrich

Neural machine learning models can successfully model language that is similar to their training distribution, but they are highly susceptible to degradation under distribution shift, which occurs in many practical applications when processing out-of-domain (OOD) text. This has been attributed to "shortcut learning": relying on weak correlations over arbitrary large contexts. We propose a method based on OOD detection with Random Network Distillation to allow an autoregressive language model to automatically disregard OOD context during inference, smoothly transitioning towards a less expressive but more robust model as the data becomes more OOD while retaining its full context capability when operating in-distribution. We apply our method to a GRU architecture, demonstrating improvements on multiple language modeling (LM) datasets.