Alignment with human preferences is an important step in developing accurate and safe large language models. This is no exception in machine translation (MT), where better handling of language nuances and context-specific variations leads to improved quality. However, preference data based on human feedback can be very expensive to obtain and curate at a large scale. Automatic metrics, on the other hand, can induce preferences, but they might not match human expectations perfectly. In this paper, we propose an approach that leverages the best of both worlds. We first collect sentence-level quality assessments from professional linguists on translations generated by multiple high-quality MT systems and evaluate the ability of current automatic metrics to recover these preferences. We then use this analysis to curate a new dataset, MT-Pref (metric induced translation preference) dataset, which comprises 18k instances covering 18 language directions, using texts sourced from multiple domains post-2022. We show that aligning TOWER models on MT-Pref significantly improves translation quality on WMT23 and FLORES benchmarks.

Hallucinated translations pose significant threats and safety concerns when it comes to practical deployment of machine translation systems. Previous research works have identified that detectors exhibit complementary performance — different detectors excel at detecting different types of hallucinations. In this paper, we propose to address the limitations of individual detectors by combining them and introducing a straightforward method for aggregating multiple detectors. Our results demonstrate the efficacy of our aggregated detector, providing a promising step towards evermore reliable machine translation systems.

Widely used learned metrics for machine translation evaluation, such as Comet and Bleurt, estimate the quality of a translation hypothesis by providing a single sentence-level score. As such, they offer little insight into translation errors (e.g., what are the errors and what is their severity). On the other hand, generative large language models (LLMs) are amplifying the adoption of more granular strategies to evaluation, attempting to detail and categorize translation errors. In this work, we introduce xcomet, an open-source learned metric designed to bridge the gap between these approaches. xcomet integrates both sentence-level evaluation and error span detection capabilities, exhibiting state-of-the-art performance across all types of evaluation (sentence-level, system-level, and error span detection). Moreover, it does so while highlighting and categorizing error spans, thus enriching the quality assessment. We also provide a robustness analysis with stress tests, and show that xcomet is largely capable of identifying localized critical errors and hallucinations.

Large language models (LLMs) have achieved state-of-the-art performance in machine translation (MT) and demonstrated the ability to leverage in-context learning through few-shot examples. However, the mechanisms by which LLMs use different parts of the input context remain largely unexplored. In this work, we provide a comprehensive analysis of context utilization in MT, studying how LLMs use various context parts, such as few-shot examples and the source text, when generating translations. We highlight several key findings: (1) the source part of few-shot examples appears to contribute more than its corresponding targets, irrespective of translation direction; (2) finetuning LLMs with parallel data alters the contribution patterns of different context parts; and (3) there is a positional bias where earlier few-shot examples have higher contributions to the translated sequence. Finally, we demonstrate that inspecting anomalous context contributions can potentially uncover pathological translations, such as hallucinations. Our findings shed light on the internal workings of LLM-based MT which go beyond those known for standard encoder-decoder MT models.

While machine translation (MT) systems are achieving increasingly strong performance on benchmarks, they often produce translations with errors and anomalies. Understanding these errors can potentially help improve the translation quality and user experience. This paper introduces xTower, an open large language model (LLM) built on top of TowerBase designed to provide free-text explanations for translation errors in order to guide the generation of a corrected translation. The quality of the generated explanations by xTower are assessed via both intrinsic and extrinsic evaluation. We ask expert translators to evaluate the quality of the explanations across two dimensions: relatedness towards the error span being explained and helpfulness in error understanding and improving translation quality. Extrinsically, we test xTower across various experimental setups in generating translation corrections, demonstrating significant improvements in translation quality. Our findings highlight xTower’s potential towards not only producing plausible and helpful explanations of automatic translations, but also leveraging them to suggest corrected translations.

We report the results of the WMT 2024 shared task on Quality Estimation, in which the challenge is to predict the quality of the output of neural machine translation systems at the word and sentence levels, without access to reference translations. In this edition, we expanded our scope to assess the potential for quality estimates to help in the correction of translated outputs, hence including an automated post-editing (APE) direction. We publish new test sets with human annotations that target two directions: providing new Multidimensional Quality Metrics (MQM) annotations for three multi-domain language pairs (English to German, Spanish and Hindi) and extending the annotations on Indic languages providing direct assessments and post edits for translation from English into Hindi, Gujarati, Tamil and Telugu. We also perform a detailed analysis of the behaviour of different models with respect to different phenomena including gender bias, idiomatic language, and numerical and entity perturbations. We received submissions based both on traditional, encoder-based approaches as well as large language model (LLM) based ones.

In this work, we present Tower v2, an improved iteration of the state-of-the-art open-weight Tower models, and the backbone of our submission to the WMT24 General Translation shared task. Tower v2 introduces key improvements including expanded language coverage, enhanced data quality, and increased model capacity up to 70B parameters. Our final submission combines these advancements with quality-aware decoding strategies, selecting translations based on multiple translation quality signals. The resulting system demonstrates significant improvement over previous versions, outperforming closed commercial systems like GPT-4o, Claude 3.5, and DeepL even at a smaller 7B scale.

Neural metrics for machine translation (MT) evaluation have become increasingly prominent due to their superior correlation with human judgments compared to traditional lexical metrics. Researchers have therefore utilized neural metrics through quality-informed decoding strategies, achieving better results than likelihood-based methods. With the rise of Large Language Models (LLMs), preference-based alignment techniques have gained attention for their potential to enhance translation quality by optimizing model weights directly on preferences induced by quality estimators. This study focuses on Contrastive Preference Optimization (CPO) and conducts extensive experiments to evaluate the impact of preference-based alignment on translation quality. Our findings indicate that while CPO consistently outperforms Supervised Fine-Tuning (SFT) on high-quality data with regard to the alignment metric, it may lead to instability across downstream evaluation metrics, particularly between neural and lexical ones. Additionally, we demonstrate that relying solely on the base model for generating candidate translations achieves performance comparable to using multiple external systems, while ensuring better consistency across downstream metrics.

Neural machine translation (NMT) has become the de-facto standard in real-world machine translation applications. However, NMT models can unpredictably produce severely pathological translations, known as hallucinations, that seriously undermine user trust. It becomes thus crucial to implement effective preventive strategies to guarantee their proper functioning. In this paper, we address the problem of hallucination detection in NMT by following a simple intuition: as hallucinations are detached from the source content, they exhibit encoder-decoder attention patterns that are statistically different from those of good quality translations. We frame this problem with an optimal transport formulation and propose a fully unsupervised, plug-in detector that can be used with any attention-based NMT model. Experimental results show that our detector not only outperforms all previous model-based detectors, but is also competitive with detectors that employ external models trained on millions of samples for related tasks such as quality estimation and cross-lingual sentence similarity.

Selective rationales and counterfactual examples have emerged as two effective, complementary classes of interpretability methods for analyzing and training NLP models. However, prior work has not explored how these methods can be integrated to combine their complementary advantages. We overcome this limitation by introducing CREST (ContRastive Edits with Sparse raTionalization), a joint framework for selective rationalization and counterfactual text generation, and show that this framework leads to improvements in counterfactual quality, model robustness, and interpretability. First, CREST generates valid counterfactuals that are more natural than those produced by previous methods, and subsequently can be used for data augmentation at scale, reducing the need for human-generated examples. Second, we introduce a new loss function that leverages CREST counterfactuals to regularize selective rationales and show that this regularization improves both model robustness and rationale quality, compared to methods that do not leverage CREST counterfactuals. Our results demonstrate that CREST successfully bridges the gap between selective rationales and counterfactual examples, addressing the limitations of existing methods and providing a more comprehensive view of a model’s predictions.

Neural metrics for machine translation evaluation, such as COMET, exhibit significant improvements in their correlation with human judgments, as compared to traditional metrics based on lexical overlap, such as BLEU. Yet, neural metrics are, to a great extent, “black boxes” returning a single sentence-level score without transparency about the decision-making process. In this work, we develop and compare several neural explainability methods and demonstrate their effectiveness for interpreting state-of-the-art fine-tuned neural metrics. Our study reveals that these metrics leverage token-level information that can be directly attributed to translation errors, as assessed through comparison of token-level neural saliency maps with Multidimensional Quality Metrics (MQM) annotations and with synthetically-generated critical translation errors. To ease future research, we release our code at: https://github.com/Unbabel/COMET/tree/explainable-metrics

Although the problem of hallucinations in neural machine translation (NMT) has received some attention, research on this highly pathological phenomenon lacks solid ground. Previous work has been limited in several ways: it often resorts to artificial settings where the problem is amplified, it disregards some (common) types of hallucinations, and it does not validate adequacy of detection heuristics. In this paper, we set foundations for the study of NMT hallucinations. First, we work in a natural setting, i.e., in-domain data without artificial noise neither in training nor in inference. Next, we annotate a dataset of over 3.4k sentences indicating different kinds of critical errors and hallucinations. Then, we turn to detection methods and both revisit methods used previously and propose using glass-box uncertainty-based detectors. Overall, we show that for preventive settings, (i) previously used methods are largely inadequate, (ii) sequence log-probability works best and performs on par with reference-based methods. Finally, we propose DeHallucinator, a simple method for alleviating hallucinations at test time that significantly reduces the hallucinatory rate.

Hallucinated translations can severely undermine and raise safety issues when machine translation systems are deployed in the wild. Previous research on the topic focused on small bilingual models trained on high-resource languages, leaving a gap in our understanding of hallucinations in multilingual models across diverse translation scenarios. In this work, we fill this gap by conducting a comprehensive analysis—over 100 language pairs across various resource levels and going beyond English-centric directions—on both the M2M neural machine translation (NMT) models and GPT large language models (LLMs). Among several insights, we highlight that models struggle with hallucinations primarily in low-resource directions and when translating out of English, where, critically, they may reveal toxic patterns that can be traced back to the training data. We also find that LLMs produce qualitatively different hallucinations to those of NMT models. Finally, we show that hallucinations are hard to reverse by merely scaling models trained with the same data. However, employing more diverse models, trained on different data or with different procedures, as fallback systems can improve translation quality and virtually eliminate certain pathologies.

We report the results of the WMT 2023 shared task on Quality Estimation, in which the challenge is to predict the quality of the output of neural machine translation systems at the word and sentence levels, without access to reference translations. This edition introduces a few novel aspects and extensions that aim to enable more fine-grained, and explainable quality estimation approaches. We introduce an updated quality annotation scheme using Multidimensional Quality Metrics to obtain sentence- and word-level quality scores for three language pairs. We also extend the provided data to new language pairs: we specifically target low-resource languages and provide training, development and test data for English-Hindi, English-Tamil, English-Telegu and English-Gujarati as well as a zero-shot test-set for English-Farsi. Further, we introduce a novel fine-grained error prediction task aspiring to motivate research towards more detailed quality predictions.

We present the joint contribution of Unbabel and Instituto Superior Técnico to the WMT 2023 Shared Task on Quality Estimation (QE). Our team participated on all tasks: Sentence- and Word-level Quality Prediction and Fine-grained error span detection. For all tasks we build on the CometKiwi model (rei et al. 2022). Our multilingual approaches are ranked first for all tasks, reaching state-of-the-art performance for quality estimation at word-, span- and sentence-level granularity. Compared to the previous state-of-the-art, CometKiwi, we show large improvements in correlation with human judgements (up to 10 Spearman points) and surpassing the second-best multilingual submission with up to 3.8 absolute points.

We present the joint contribution of IST and Unbabel to the WMT 2022 Shared Task on Quality Estimation (QE). Our team participated in all three subtasks: (i) Sentence and Word-level Quality Prediction; (ii) Explainable QE; and (iii) Critical Error Detection. For all tasks we build on top of the COMET framework, connecting it with the predictor-estimator architecture of OpenKiwi, and equipping it with a word-level sequence tagger and an explanation extractor. Our results suggest that incorporating references during pretraining improves performance across several language pairs on downstream tasks, and that jointly training with sentence and word-level objectives yields a further boost. Furthermore, combining attention and gradient information proved to be the top strategy for extracting good explanations of sentence-level QE models. Overall, our submissions achieved the best results for all three tasks for almost all language pairs by a considerable margin.

We present the joint contribution of Instituto Superior Técnico (IST) and Unbabel to the Explainable Quality Estimation (QE) shared task, where systems were submitted to two tracks: constrained (without word-level supervision) and unconstrained (with word-level supervision). For the constrained track, we experimented with several explainability methods to extract the relevance of input tokens from sentence-level QE models built on top of multilingual pre-trained transformers. Among the different tested methods, composing explanations in the form of attention weights scaled by the norm of value vectors yielded the best results. When word-level labels are used during training, our best results were obtained by using word-level predicted probabilities. We further improve the performance of our methods on the two tracks by ensembling explanation scores extracted from models trained with different pre-trained transformers, achieving strong results for in-domain and zero-shot language pairs.

Selective rationalization aims to produce decisions along with rationales (e.g., text highlights or word alignments between two sentences). Commonly, rationales are modeled as stochastic binary masks, requiring sampling-based gradient estimators, which complicates training and requires careful hyperparameter tuning. Sparse attention mechanisms are a deterministic alternative, but they lack a way to regularize the rationale extraction (e.g., to control the sparsity of a text highlight or the number of alignments). In this paper, we present a unified framework for deterministic extraction of structured explanations via constrained inference on a factor graph, forming a differentiable layer. Our approach greatly eases training and rationale regularization, generally outperforming previous work on what comes to performance and plausibility of the extracted rationales. We further provide a comparative study of stochastic and deterministic methods for rationale extraction for classification and natural language inference tasks, jointly assessing their predictive power, quality of the explanations, and model variability.