Neural Machine Translation models tend to perpetuate gender bias present in their training data distribution. Context-aware models have been previously suggested as a means to mitigate this type of bias. In this work, we examine this claim by analysing in detail the translation of stereotypical professions in English to German, and translation with non-informative context in Basque to Spanish. Our results show that, although context-aware models can significantly enhance translation accuracy for feminine terms, they can still maintain or even amplify gender bias. These results highlight the need for more fine-grained approaches to bias mitigation in Neural Machine Translation.

Standard context-aware neural machine translation (NMT) typically relies on parallel document-level data, exploiting both source and target contexts. Concatenation-based approaches in particular, still a strong baseline for document-level NMT, prepend source and/or target context sentences to the sentences to be translated, with model variants that exploit equal amounts of source and target data on each side achieving state-of-the-art results. In this work, we investigate whether target data should be further promoted within standard concatenation-based approaches, as most document-level phenomena rely on information that is present on the target language side. We evaluate novel concatenation-based variants where the target context is prepended to the source language, either in isolation or in combination with the source context. Experimental results in English-Russian and Basque-Spanish show that including target context in the source leads to large improvements on target language phenomena. On source-dependent phenomena, using only target language context in the source achieves parity with state-of-the-art concatenation approaches, or slightly underperforms, whereas combining source and target context on the source side leads to significant gains across the board.

We describe Vicomtech’s participation in the WMT 2024 Shared Task on translation into low-resource languages of Spain. We addressed all three languages of the task, namely Aragonese, Aranese and Asturian, in both constrained and open settings. Our work mainly centred on exploiting different types of corpora via data filtering, selection and combination methods, along with synthetic data generated with translation models based on rules, neural sequence-to-sequence or large language models. We improved or matched the best baselines in all three language pairs and present complementary results on additional test sets.

The Split and Rephrase (SPRP) task, which consists in splitting complex sentences into a sequence of shorter grammatical sentences, while preserving the original meaning, can facilitate the processing of complex texts for humans and machines alike. It is also a valuable testbed to evaluate natural language processing models, as it requires modelling complex grammatical aspects. In this work, we evaluate large language models on the task, showing that they can provide large improvements over the state of the art on the main metrics, although still lagging in terms of splitting compliance. Results from two human evaluations further support the conclusions drawn from automated metric results. We provide a comprehensive study that includes prompting variants, domain shift, fine-tuned pretrained language models of varying parameter size and training data volumes, contrasted with both zero-shot and few-shot approaches on instruction-tuned language models. Although the latter were markedly outperformed by fine-tuned models, they may constitute a reasonable off-the-shelf alternative. Our results provide a fine-grained analysis of the potential and limitations of large language models for SPRP, with significant improvements achievable using relatively small amounts of training data and model parameters overall, and remaining limitations for all models on the task.

Document-level Machine Translation has emerged as a promising means to enhance automated translation quality, but it is currently unclear how effectively context-aware models use the available context during translation. This paper aims to provide insight into the current state of models based on input concatenation, with an in-depth evaluation on English–German and English–French standard datasets. We notably evaluate the impact of data bias, antecedent part-of-speech, context complexity, and the syntactic function of the elements involved in discursive phenomena. Our experimental results indicate that the selected models do improve the overall translation in context, with varying sensitivity to the different factors we examined. We notably show that the selected context-aware models operate markedly better on regular syntactic configurations involving subject antecedents and pronouns, with degraded performance as the configurations become more dissimilar.

We explore the use of source factors in context-aware neural machine translation, specifically concatenation-based models, to improve the translation quality of inter-sentential phenomena. Context sentences are typically concatenated to the sentence to be translated, with string-based markers to separate the latter from the former. Although previous studies have measured the impact of prefixes to identify and mark context information, the use of learnable factors has only been marginally explored. In this study, we evaluate the impact of single and multiple source context factors in English-German and Basque-Spanish contextual translation. We show that this type of factors can significantly enhance translation accuracy for phenomena such as gender and register coherence in Basque-Spanish, while also improving BLEU results in some scenarios. These results demonstrate the potential of factor-based context identification to improve context-aware machine translation in future research.

Progress in document-level Machine Translation is hindered by the lack of parallel training data that include context information. In this work, we evaluate the potential of data augmentation techniques to circumvent these limitations, showing that significant gains can be achieved via upsampling, similar context sampling and back-translations, targeted on context-relevant data. We apply these methods on standard document-level datasets in English-German and English-French and demonstrate their relevance to improve the translation of contextual phenomena. In particular, we show that relatively small volumes of targeted data augmentation lead to significant improvements over a strong context-concatenation baseline and standard back-translation of document-level data. We also compare the accuracy of the selected methods depending on data volumes or distance to relevant context information, and explore their use in combination.

Document-level Neural Machine Translation aims to increase the quality of neural translation models by taking into account contextual information. Properly modelling information beyond the sentence level can result in improved machine translation output in terms of coherence, cohesion and consistency. Suitable corpora for context-level modelling are necessary to both train and evaluate context-aware systems, but are still relatively scarce. In this work we describe TANDO, a document-level corpus for the under-resourced Basque-Spanish language pair, which we share with the scientific community. The corpus is composed of parallel data from three different domains and has been prepared with context-level information. Additionally, the corpus includes contrastive test sets for fine-grained evaluations of gender and register contextual phenomena on both source and target language sides. To establish the usefulness of the corpus, we trained and evaluated baseline Transformer models and context-aware variants based on context concatenation. Our results indicate that the corpus is suitable for fine-grained evaluation of document-level machine translation systems.

Adaptive Machine Translation purports to dynamically include user feedback to improve translation quality. In a post-editing scenario, user corrections of machine translation output are thus continuously incorporated into translation models, reducing or eliminating repetitive error editing and increasing the usefulness of automated translation. In neural machine translation, this goal may be achieved via online learning approaches, where network parameters are updated based on each new sample. This type of adaptation typically requires higher learning rates, which can affect the quality of the models over time. Alternatively, less aggressive online learning setups may preserve model stability, at the cost of reduced adaptation to user-generated corrections. In this work, we evaluate different online learning configurations over time, measuring their impact on user-generated samples, as well as separate in-domain and out-of-domain datasets. Results in two different domains indicate that mixed approaches combining online learning with periodic batch fine-tuning might be needed to balance the benefits of online learning with model stability.

We present a comparative evaluation of casing methods for Neural Machine Translation, to help establish an optimal pre- and post-processing methodology. We trained and compared system variants on data prepared with the main casing methods available, namely translation of raw data without case normalisation, lowercasing with recasing, truecasing, case factors and inline casing. Machine translation models were prepared on WMT 2017 English-German and English-Turkish datasets, for all translation directions, and the evaluation includes reference metric results as well as a targeted analysis of case preservation accuracy. Inline casing, where case information is marked along lowercased words in the training data, proved to be the optimal approach overall in these experiments.

We present the results of a case study in the exploitation of comparable corpora for Neural Machine Translation. A large comparable corpus for Basque-Spanish was prepared, on the basis of independently-produced news by the Basque public broadcaster EiTB, and we discuss the impact of various techniques to exploit the original data in order to determine optimal variants of the corpus. In particular, we show that filtering in terms of alignment thresholds and length-difference outliers has a significant impact on translation quality. The impact of tags identifying comparable data in the training datasets is also evaluated, with results indicating that this technique might be useful to help the models discriminate noisy information, in the form of informational imbalance between aligned sentences. The final corpus was prepared according to the experimental results and is made available to the scientific community for research purposes.