Large language models (LLMs) are increasingly popular but are also prone to generating bias, toxic or harmful language, which can have detrimental effects on individuals and communities. Although most efforts is put to assess and mitigate toxicity in generated content, it is primarily concentrated on English, while it’s essential to consider other languages as well. For addressing this issue, we create and release FrenchToxicityPrompts, a dataset of 50K naturally occurring French prompts and their continuations, annotated with toxicity scores from a widely used toxicity classifier. We evaluate 14 different models from four prevalent open-sourced families of LLMs against our dataset to assess their potential toxicity across various dimensions. We hope that our contribution will foster future research on toxicity detection and mitigation beyond English.

Retrieval-augmented generation (RAG) has recently emerged as a promising solution for incorporating up-to-date or domain-specific knowledge into large language models (LLMs) and improving LLM factuality, but is predominantly studied in English-only settings. In this work, we consider RAG in the multilingual setting (mRAG), i.e. with user queries and the datastore in 13 languages, and investigate which components and with which adjustments are needed to build a well-performing mRAG pipeline, that can be used as a strong baseline in future works. Our findings highlight that despite the availability of high-quality off-the-shelf multilingual retrievers and generators, task-specific prompt engineering is needed to enable generation in user languages. Moreover, current evaluation metrics need adjustments for multilingual setting, to account for variations in spelling named entities. The main limitations to be addressed in future works include frequent code-switching in non-Latin alphabet languages, occasional fluency errors, wrong reading of the provided documents, or irrelevant retrieval. We release the code for the resulting mRAG baseline pipeline at https://github.com/naver/bergen, Documentation: https://github.com/naver/bergen/blob/main/documentations/multilingual.md.

Retrieval-Augmented Generation allows to enhance Large Language Models with external knowledge. In response to the recent popularity of generative LLMs, many RAG approaches have been proposed, which involve an intricate number of different configurations such as evaluation datasets, collections, metrics, retrievers, and LLMs. Inconsistent benchmarking poses a major challenge in comparing approaches and understanding the impact of each component in the pipeline. In this work, we study best practices that lay the groundwork for a systematic evaluation of RAG and present BERGEN, an end-to-end library for reproducible research standardizing RAG experiments. In an extensive study focusing on QA, we benchmark different state-of-the-art retrievers, rerankers, and LLMs. Additionally, we analyze existing RAG metrics and datasets.

Instruction tuning (IT) is widely used to teach pretrained large language models (LLMs) to follow arbitrary instructions, but is under-studied in multilingual settings. In this work, we conduct a systematic study of zero-shot cross-lingual transfer in IT, when an LLM is instruction-tuned on English-only data and then tested on user prompts in other languages. We advocate for the importance of evaluating various aspects of model responses in multilingual instruction following and investigate the influence of different model configuration choices. We find that cross-lingual transfer does happen successfully in IT even if all stages of model training are English-centric, but only if multiliguality is taken into account in hyperparameter tuning and with large enough IT data. English-trained LLMs are capable of generating correct-language, comprehensive and helpful responses in other languages, but suffer from low factuality and may occasionally have fluency errors.

Zero-shot cross-lingual transfer, which implies finetuning of the multilingual pretrained language model on input-output pairs in one language and using it to make task predictions for inputs in other languages, was widely studied for natural language understanding but is understudied for generation. Previous works notice a frequent problem of generation in a wrong language and propose approaches to address it, usually using mT5 as a backbone model. In this work we compare various approaches proposed from the literature in unified settings, also including alternative backbone models, namely mBART and NLLB-200. We first underline the importance of tuning learning rate used for finetuning, which helps to substantially alleviate the problem of generation in the wrong language. Then, we show that with careful learning rate tuning, the simple full finetuning of the model acts as a very strong baseline and alternative approaches bring only marginal improvements. Finally, we find that mBART performs similarly to mT5 of the same size, and NLLB-200 can be competitive in some cases. Our final zero-shot models reach the performance of the approach based on data translation which is usually considered as an upper baseline for zero-shot cross-lingual transfer in generation.

The recently released NLLB-200 is a set of multilingual Neural Machine Translation models that cover 202 languages. The largest model is based on a Mixture of Experts architecture and achieves SoTA results across many language pairs. It contains 54.5B parameters and requires at least four 32GB GPUs just for inference. In this work, we propose a pruning method that enables the removal of up to 80% of experts without further finetuning and with a negligible loss in translation quality, which makes it feasible to run the model on a single 32GB GPU. Further analysis suggests that our pruning metrics can identify language-specific experts.

The BLOOM model is a large publicly available multilingual language model, but its pretraining was limited to 46 languages. To extend the benefits of BLOOM to other languages without incurring prohibitively large costs, it is desirable to adapt BLOOM to new languages not seen during pretraining. In this work, we apply existing language adaptation strategies to BLOOM and benchmark its zero-shot prompting performance on eight new languages in a resource-constrained setting. We find language adaptation to be effective at improving zero-shot performance in new languages. Surprisingly, we find that adapter-based finetuning is more effective than continued pretraining for large models. In addition, we discover that prompting performance is not significantly affected by language specifics, such as the writing system. It is primarily determined by the size of the language adaptation data. We also add new languages to BLOOMZ, which is a multitask finetuned version of BLOOM capable of following task instructions zero-shot. We find including a new language in the multitask fine-tuning mixture to be the most effective method to teach BLOOMZ a new language. We conclude that with sufficient training data language adaptation can generalize well to diverse languages. Our code is available at https://github.com/bigscience-workshop/multilingual-modeling.

We study the zero-shot setting for the aspect-based scientific document summarization task. Summarizing scientific documents with respect to an aspect can remarkably improve document assistance systems and readers experience. However, existing large-scale datasets contain a limited variety of aspects, causing summarization models to over-fit to a small set of aspects and a specific domain. We establish baseline results in zero-shot performance (over unseen aspects and the presence of domain shift), paraphrasing, leave-one-out, and limited supervised samples experimental setups. We propose a self-supervised pre-training approach to enhance the zero-shot performance. We leverage the PubMed structured abstracts to create a biomedical aspect-based summarization dataset. Experimental results on the PubMed and FacetSum aspect-based datasets show promising performance when the model is pre-trained using unlabelled in-domain data.

In recent years, multilingual machine translation models have achieved promising performance on low-resource language pairs by sharing information between similar languages, thus enabling zero-shot translation. To overcome the “curse of multilinguality”, these models often opt for scaling up the number of parameters, which makes their use in resource-constrained environments challenging. We introduce SMaLL-100, a distilled version of the M2M-100(12B) model, a massively multilingual machine translation model covering 100 languages. We train SMaLL-100 with uniform sampling across all language pairs and therefore focus on preserving the performance of low-resource languages. We evaluate SMaLL-100 on different low-resource benchmarks: FLORES-101, Tatoeba, and TICO-19 and demonstrate that it outperforms previous massively multilingual models of comparable sizes (200-600M) while improving inference latency and memory usage. Additionally, our model achieves comparable results to M2M-100 (1.2B), while being 3.6x smaller and 4.3x faster at inference.

Domain Adaptation (DA) of Neural Machine Translation (NMT) model often relies on a pre-trained general NMT model which is adapted to the new domain on a sample of in-domain parallel data. Without parallel data, there is no way to estimate the potential benefit of DA, nor the amount of parallel samples it would require. It is however a desirable functionality that could help MT practitioners to make an informed decision before investing resources in dataset creation. We propose a Domain adaptation Learning Curve prediction (DaLC) model that predicts prospective DA performance based on in-domain monolingual samples in the source language. Our model relies on the NMT encoder representations combined with various instance and corpus-level features. We demonstrate that instance-level is better able to distinguish between different domains compared to corpus-level frameworks proposed in previous studies Finally, we perform in-depth analyses of the results highlighting the limitations of our approach, and provide directions for future research.

Softmax is the de facto standard for normalizing logits in modern neural networks for language processing. However, by producing a dense probability distribution each token in the vocabulary has a nonzero chance of being selected at each generation step, leading to a variety of reported problems in text generation. 𝛼-entmax of Peters et al. (2019) solves this problem, but is unfortunately slower than softmax. In this paper, we propose an alternative to 𝛼-entmax, which keeps its virtuous characteristics, but is as fast as optimized softmax and achieves on par or better performance in machine translation task.

Recently, very large pre-trained models achieve state-of-the-art results in various natural language processing (NLP) tasks, but their size makes it more challenging to apply them in resource-constrained environments. Compression techniques allow to drastically reduce the size of the models and therefore their inference time with negligible impact on top-tier metrics. However, the general performance averaged across multiple tasks and/or languages may hide a drastic performance drop on under-represented features, which could result in the amplification of biases encoded by the models. In this work, we assess the impact of compression methods on Multilingual Neural Machine Translation models (MNMT) for various language groups, gender, and semantic biases by extensive analysis of compressed models on different machine translation benchmarks, i.e. FLORES-101, MT-Gender, and DiBiMT. We show that the performance of under-represented languages drops significantly, while the average BLEU metric only slightly decreases. Interestingly, the removal of noisy memorization with compression leads to a significant improvement for some medium-resource languages. Finally, we demonstrate that compression amplifies intrinsic gender and semantic biases, even in high-resource languages.

This paper presents an interactive data dashboard that provides users with an overview of the preservation of discourse relations among 28 language pairs. We display a graph network depicting the cross-lingual discourse relations between a pair of languages for multilingual TED talks and provide a search function to look for sentences with specific keywords or relation types, facilitating ease of analysis on the cross-lingual discourse relations.

Adapter layers are lightweight, learnable units inserted between transformer layers. Recent work explores using such layers for neural machine translation (NMT), to adapt pre-trained models to new domains or language pairs, training only a small set of parameters for each new setting (language pair or domain). In this work we study the compositionality of language and domain adapters in the context of Machine Translation. We aim to study, 1) parameter-efficient adaptation to multiple domains and languages simultaneously (full-resource scenario) and 2) cross-lingual transfer in domains where parallel data is unavailable for certain language pairs (partial-resource scenario). We find that in the partial resource scenario a naive combination of domain-specific and language-specific adapters often results in ‘catastrophic forgetting’ of the missing languages. We study other ways to combine the adapters to alleviate this issue and maximize cross-lingual transfer. With our best adapter combinations, we obtain improvements of 3-4 BLEU on average for source languages that do not have in-domain data. For target languages without in-domain data, we achieve a similar improvement by combining adapters with back-translation. Supplementary material is available at https://tinyurl.com/r66stbxj.

Language domains that require very careful use of terminology are abundant and reflect a significant part of the translation industry. In this work we introduce a benchmark for evaluating the quality and consistency of terminology translation, focusing on the medical (and COVID-19 specifically) domain for five language pairs: English to French, Chinese, Russian, and Korean, as well as Czech to German. We report the descriptions and results of the participating systems, commenting on the need for further research efforts towards both more adequate handling of terminologies as well as towards a proper formulation and evaluation of the task.

Multilingual NMT has become an attractive solution for MT deployment in production. But to match bilingual quality, it comes at the cost of larger and slower models. In this work, we consider several ways to make multilingual NMT faster at inference without degrading its quality. We experiment with several “light decoder” architectures in two 20-language multi-parallel settings: small-scale on TED Talks and large-scale on ParaCrawl. Our experiments demonstrate that combining a shallow decoder with vocabulary filtering leads to almost 2 times faster inference with no loss in translation quality. We validate our findings with BLEU and chrF (on 380 language pairs), robustness evaluation and human evaluation.

This paper describes Naver Labs Europe’s participation in the Robustness, Chat, and Biomedical Translation tasks at WMT 2020. We propose a bidirectional German-English model that is multi-domain, robust to noise, and which can translate entire documents (or bilingual dialogues) at once. We use the same ensemble of such models as our primary submission to all three tasks and achieve competitive results. We also experiment with language model pre-training techniques and evaluate their impact on robustness to noise and out-of-domain translation. For German, Spanish, Italian, and French to English translation in the Biomedical Task, we also submit our recently released multilingual Covid19NMT model.

We release a multilingual neural machine translation model, which can be used to translate text in the biomedical domain. The model can translate from 5 languages (French, German, Italian, Korean and Spanish) into English. It is trained with large amounts of generic and biomedical data, using domain tags. Our benchmarks show that it performs near state-of-the-art both on news (generic domain) and biomedical test sets, and that it outperforms the existing publicly released models. We believe that this release will help the large-scale multilingual analysis of the digital content of the COVID-19 crisis and of its effects on society, economy, and healthcare policies. We also release a test set of biomedical text for Korean-English. It consists of 758 sentences from official guidelines and recent papers, all about COVID-19.

Exploiting large pretrained models for various NMT tasks have gained a lot of visibility recently. In this work we study how BERT pretrained models could be exploited for supervised Neural Machine Translation. We compare various ways to integrate pretrained BERT model with NMT model and study the impact of the monolingual data used for BERT training on the final translation quality. We use WMT-14 English-German, IWSLT15 English-German and IWSLT14 English-Russian datasets for these experiments. In addition to standard task test set evaluation, we perform evaluation on out-of-domain test sets and noise injected test sets, in order to assess how BERT pretrained representations affect model robustness.

We share a French-English parallel corpus of Foursquare restaurant reviews, and define a new task to encourage research on Neural Machine Translation robustness and domain adaptation, in a real-world scenario where better-quality MT would be greatly beneficial. We discuss the challenges of such user-generated content, and train good baseline models that build upon the latest techniques for MT robustness. We also perform an extensive evaluation (automatic and human) that shows significant improvements over existing online systems. Finally, we propose task-specific metrics based on sentiment analysis or translation accuracy of domain-specific polysemous words.

In this paper, we test state-of-the-art Aspect Based Sentiment Analysis (ABSA) systems trained on a widely used dataset on actual data. We created a new manually annotated dataset of user generated data from the same domain as the training dataset, but from other sources and analyse the differences between the new and the standard ABSA dataset. We then analyse the results in performance of different versions of the same system on both datasets. We also propose light adaptation methods to increase system robustness.