Multilingual parallel data for speech-to-speech translation is scarce and expensive to create from scratch. This is all the more true for expressive speech translation, which aims at preserving not only the semantics, but also the overall prosody (e.g. style, emotion, rate-of-speech). Existing corpora contain speech utterances with the same meaning, yet the overall prosody is typically different, as human annotators are not tasked with reproducing these aspects, or crowed-sourced efforts do not specifically target this kind of alignment in priority. In this paper, we propose a novel alignment algorithm, which automatically forms pairs of speech segments aligned not only in meaning, but also in expressivity. In order to validate our approach, we train an expressive multilingual speech-to-speech translation system on the automatically aligned data. Our experiments show that in comparison to semantic-only approaches, expressively aligned data yields large improvements in source expressivity preservation (e.g. 43% uplift in speech rate preservation on average), while still maintaining content translation quality. In some scenarios, results also indicate that this alignment algorithm can outperform standard, semantic-focused approaches even on content translation quality.

End-to-End speech-to-speech translation (S2ST) is generally evaluated with text-based metrics. This means that generated speech has to be automatically transcribed, making the evaluation dependent on the availability and quality of automatic speech recognition (ASR) systems. In this paper, we propose a text-free evaluation metric for end-to-end S2ST, named BLASER, to avoid the dependency on ASR systems. BLASER leverages a multilingual multimodal encoder to directly encode the speech segments for source input, translation output and reference into a shared embedding space and computes a score of the translation quality that can be used as a proxy to human evaluation. To evaluate our approach, we construct training and evaluation sets from more than 40k human annotations covering seven language directions. The best results of BLASER are achieved by training with supervision from human rating scores. We show that when evaluated at the sentence level, BLASER correlates significantly better with human judgment compared to ASR dependent metrics including ASR-SENTBLEU in all translation directions and ASR-COMET in five of them. Our analysis shows combining speech and text as inputs to BLASER does not increase the correlation with human scores, but best correlations are achieved when using speech, which motivates the goal of our research. Moreover, we show that using ASR for references is detrimental for text-based metrics.

We present SpeechMatrix, a large-scale multilingual corpus of speech-to-speech translations mined from real speech of European Parliament recordings. It contains speech alignments in 136 language pairs with a total of 418 thousand hours of speech. To evaluate the quality of this parallel speech, we train bilingual speech-to-speech translation models on mined data only and establish extensive baseline results on EuroParl-ST, VoxPopuli and FLEURS test sets. Enabled by the multilinguality of SpeechMatrix, we also explore multilingual speech-to-speech translation, a topic which was addressed by few other works. We also demonstrate that model pre-training and sparse scaling using Mixture-of-Experts bring large gains to translation performance. The mined data and models will be publicly released

We introduce a new proxy score for evaluating bitext mining based on similarity in a multilingual embedding space: xsim++. In comparison to xsim, this improved proxy leverages rule-based approaches to extend English sentences in any evaluation set with synthetic, hard-to-distinguish examples which more closely mirror the scenarios we encounter during large-scale mining. We validate this proxy by running a significant number of bitext mining experiments for a set of low-resource languages, and subsequently train NMT systems on the mined data. In comparison to xsim, we show that xsim++ is better correlated with the downstream BLEU scores of translation systems trained on mined bitexts, providing a reliable proxy of bitext mining performance without needing to run expensive bitext mining pipelines. xsim++ also reports performance for different error types, offering more fine-grained feedbacks for model development.

We study speech-to-speech translation (S2ST) that translates speech from one language into another language and focuses on building systems to support languages without standard text writing systems. We use English-Taiwanese Hokkien as a case study, and present an end-to-end solution from training data collection, modeling choices to benchmark dataset release. First, we present efforts on creating human annotated data, automatically mining data from large unlabeled speech datasets, and adopting pseudo-labeling to produce weakly supervised data. On the modeling, we take advantage of recent advances in applying self-supervised discrete representations as target for prediction in S2ST and show the effectiveness of leveraging additional text supervision from Mandarin, a language similar to Hokkien, in model training. Finally, we release an S2ST benchmark set to facilitate future research in this field.

Multilingual sentence representations from large models encode semantic information from two or more languages and can be used for different cross-lingual information retrieval and matching tasks. In this paper, we integrate contrastive learning into multilingual representation distillation and use it for quality estimation of parallel sentences (i.e., find semantically similar sentences that can be used as translations of each other). We validate our approach with multilingual similarity search and corpus filtering tasks. Experiments across different low-resource languages show that our method greatly outperforms previous sentence encoders such as LASER, LASER3, and LaBSE.

We present a new approach to perform zero-shot cross-modal transfer between speech and text for translation tasks. Multilingual speech and text are encoded in a joint fixed-size representation space. Then, we compare different approaches to decode these multimodal and multilingual fixed-size representations, enabling zero-shot translation between languages and modalities. All our models are trained without the need of cross-modal labeled translation data.Despite a fixed-size representation, we achieve very competitive results on several text and speech translation tasks. In particular, we significantly improve the state-of-the-art for zero-shot speech translation on Must-C. Incorporating a speech decoder in our framework, we introduce the first results for zero-shot direct speech-to-speech and text-to-speech translation.

Neural machine translation, as other natural language deep learning applications, is hungry for data. As research evolves, the data pipelines supporting that research evolve too, oftentimes re-implementing the same core components. Despite the potential of modular codebases, researchers have but little time to put code structure and reusability first. Unfortunately, this makes it very hard to publish clean, reproducible code to benefit a wider audience. In this paper, we motivate and describe stopes , a framework that addresses these issues while empowering scalability and versatility for research use cases. This library was a key enabler of the No Language Left Behind project, establishing new state of the art performance for a multilingual machine translation model covering 200 languages. stopes and the pipelines described are released under the MIT license at https://github.com/facebookresearch/stopes.

We present a textless speech-to-speech translation (S2ST) system that can translate speech from one language into another language and can be built without the need of any text data. Different from existing work in the literature, we tackle the challenge in modeling multi-speaker target speech and train the systems with real-world S2ST data. The key to our approach is a self-supervised unit-based speech normalization technique, which finetunes a pre-trained speech encoder with paired audios from multiple speakers and a single reference speaker to reduce the variations due to accents, while preserving the lexical content. With only 10 minutes of paired data for speech normalization, we obtain on average 3.2 BLEU gain when training the S2ST model on the VoxPopuli S2ST dataset, compared to a baseline trained on un-normalized speech target. We also incorporate automatically mined S2ST data and show an additional 2.0 BLEU gain. To our knowledge, we are the first to establish a textless S2ST technique that can be trained with real-world data and works for multiple language pairs.

We present the results of the WMT’22 SharedTask on Large-Scale Machine Translation Evaluation for African Languages. The shared taskincluded both a data and a systems track, alongwith additional innovations, such as a focus onAfrican languages and extensive human evaluation of submitted systems. We received 14system submissions from 8 teams, as well as6 data track contributions. We report a largeprogress in the quality of translation for Africanlanguages since the last iteration of this sharedtask: there is an increase of about 7.5 BLEUpoints across 72 language pairs, and the average BLEU scores went from 15.09 to 22.60.

Scaling multilingual representation learning beyond the hundred most frequent languages is challenging, in particular to cover the long tail of low-resource languages. We move away from the popular one-for-all multilingual models and focus on training multiple language (family) specific representations, but most prominently enable all languages to still be encoded in the same representational space. We focus on teacher-student training, allowing all encoders to be mutually compatible for bitext mining, and enabling fast learning of new languages. We also combine supervised and self-supervised training, allowing encoders to take advantage of monolingual training data.Our approach significantly outperforms the original LASER encoder. We study very low-resource languages and handle 44 African languages, many of which are not covered by any other model. For these languages, we train sentence encoders and mine bitexts. Adding these mined bitexts yielded an improvement of 3.8 BLEU for NMT into English.

We show that margin-based bitext mining in a multilingual sentence space can be successfully scaled to operate on monolingual corpora of billions of sentences. We use 32 snapshots of a curated common crawl corpus (Wenzel et al, 2019) totaling 71 billion unique sentences. Using one unified approach for 90 languages, we were able to mine 10.8 billion parallel sentences, out of which only 2.9 billions are aligned with English. We illustrate the capability of our scalable mining system to create high quality training sets from one language to any other by training hundreds of different machine translation models and evaluating them on the many-to-many TED benchmark. Further, we evaluate on competitive translation benchmarks such as WMT and WAT. Using only mined bitext, we set a new state of the art for a single system on the WMT’19 test set for English-German/Russian/Chinese. In particular, our English/German and English/Russian systems outperform the best single ones by over 4 BLEU points and are on par with best WMT’19 systems, which train on the WMT training data and augment it with backtranslation. We also achieve excellent results for distant languages pairs like Russian/Japanese, outperforming the best submission at the 2020 WAT workshop. All of the mined bitext will be freely available.

We present an approach based on multilingual sentence embeddings to automatically extract parallel sentences from the content of Wikipedia articles in 96 languages, including several dialects or low-resource languages. We do not limit the extraction process to alignments with English, but we systematically consider all possible language pairs. In total, we are able to extract 135M parallel sentences for 16720 different language pairs, out of which only 34M are aligned with English. This corpus is freely available. To get an indication on the quality of the extracted bitexts, we train neural MT baseline systems on the mined data only for 1886 languages pairs, and evaluate them on the TED corpus, achieving strong BLEU scores for many language pairs. The WikiMatrix bitexts seem to be particularly interesting to train MT systems between distant languages without the need to pivot through English.

In this paper, we describe our end-to-end multilingual speech translation system submitted to the IWSLT 2021 evaluation campaign on the Multilingual Speech Translation shared task. Our system is built by leveraging transfer learning across modalities, tasks and languages. First, we leverage general-purpose multilingual modules pretrained with large amounts of unlabelled and labelled data. We further enable knowledge transfer from the text task to the speech task by training two tasks jointly. Finally, our multilingual model is finetuned on speech translation task-specific data to achieve the best translation results. Experimental results show our system outperforms the reported systems, including both end-to-end and cascaded based approaches, by a large margin. In some translation directions, our speech translation results evaluated on the public Multilingual TEDx test set are even comparable with the ones from a strong text-to-text translation system, which uses the oracle speech transcripts as input.

Question answering (QA) models have shown rapid progress enabled by the availability of large, high-quality benchmark datasets. Such annotated datasets are difficult and costly to collect, and rarely exist in languages other than English, making building QA systems that work well in other languages challenging. In order to develop such systems, it is crucial to invest in high quality multilingual evaluation benchmarks to measure progress. We present MLQA, a multi-way aligned extractive QA evaluation benchmark intended to spur research in this area. MLQA contains QA instances in 7 languages, English, Arabic, German, Spanish, Hindi, Vietnamese and Simplified Chinese. MLQA has over 12K instances in English and 5K in each other language, with each instance parallel between 4 languages on average. We evaluate state-of-the-art cross-lingual models and machine-translation-based baselines on MLQA. In all cases, transfer results are shown to be significantly behind training-language performance.

In this paper, we describe our submission to the WMT19 low-resource parallel corpus filtering shared task. Our main approach is based on the LASER toolkit (Language-Agnostic SEntence Representations), which uses an encoder-decoder architecture trained on a parallel corpus to obtain multilingual sentence representations. We then use the representations directly to score and filter the noisy parallel sentences without additionally training a scoring function. We contrast our approach to other promising methods and show that LASER yields strong results. Finally, we produce an ensemble of different scoring methods and obtain additional gains. Our submission achieved the best overall performance for both the Nepali-English and Sinhala-English 1M tasks by a margin of 1.3 and 1.4 BLEU respectively, as compared to the second best systems. Moreover, our experiments show that this technique is promising for low and even no-resource scenarios.

We introduce an architecture to learn joint multilingual sentence representations for 93 languages, belonging to more than 30 different families and written in 28 different scripts. Our system uses a single BiLSTM encoder with a shared byte-pair encoding vocabulary for all languages, which is coupled with an auxiliary decoder and trained on publicly available parallel corpora. This enables us to learn a classifier on top of the resulting embeddings using English annotated data only, and transfer it to any of the 93 languages without any modification. Our experiments in cross-lingual natural language inference (XNLI data set), cross-lingual document classification (MLDoc data set), and parallel corpus mining (BUCC data set) show the effectiveness of our approach. We also introduce a new test set of aligned sentences in 112 languages, and show that our sentence embeddings obtain strong results in multilingual similarity search even for low- resource languages. Our implementation, the pre-trained encoder, and the multilingual test set are available at https://github.com/facebookresearch/LASER.

Machine translation is highly sensitive to the size and quality of the training data, which has led to an increasing interest in collecting and filtering large parallel corpora. In this paper, we propose a new method for this task based on multilingual sentence embeddings. In contrast to previous approaches, which rely on nearest neighbor retrieval with a hard threshold over cosine similarity, our proposed method accounts for the scale inconsistencies of this measure, considering the margin between a given sentence pair and its closest candidates instead. Our experiments show large improvements over existing methods. We outperform the best published results on the BUCC mining task and the UN reconstruction task by more than 10 F1 and 30 precision points, respectively. Filtering the English-German ParaCrawl corpus with our approach, we obtain 31.2 BLEU points on newstest2014, an improvement of more than one point over the best official filtered version.

State-of-the-art natural language processing systems rely on supervision in the form of annotated data to learn competent models. These models are generally trained on data in a single language (usually English), and cannot be directly used beyond that language. Since collecting data in every language is not realistic, there has been a growing interest in cross-lingual language understanding (XLU) and low-resource cross-language transfer. In this work, we construct an evaluation set for XLU by extending the development and test sets of the Multi-Genre Natural Language Inference Corpus (MultiNLI) to 14 languages, including low-resource languages such as Swahili and Urdu. We hope that our dataset, dubbed XNLI, will catalyze research in cross-lingual sentence understanding by providing an informative standard evaluation task. In addition, we provide several baselines for multilingual sentence understanding, including two based on machine translation systems, and two that use parallel data to train aligned multilingual bag-of-words and LSTM encoders. We find that XNLI represents a practical and challenging evaluation suite, and that directly translating the test data yields the best performance among available baselines.

We learn a joint multilingual sentence embedding and use the distance between sentences in different languages to filter noisy parallel data and to mine for parallel data in large news collections. We are able to improve a competitive baseline on the WMT’14 English to German task by 0.3 BLEU by filtering out 25% of the training data. The same approach is used to mine additional bitexts for the WMT’14 system and to obtain competitive results on the BUCC shared task to identify parallel sentences in comparable corpora. The approach is generic, it can be applied to many language pairs and it is independent of the architecture of the machine translation system.

Many modern NLP systems rely on word embeddings, previously trained in an unsupervised manner on large corpora, as base features. Efforts to obtain embeddings for larger chunks of text, such as sentences, have however not been so successful. Several attempts at learning unsupervised representations of sentences have not reached satisfactory enough performance to be widely adopted. In this paper, we show how universal sentence representations trained using the supervised data of the Stanford Natural Language Inference datasets can consistently outperform unsupervised methods like SkipThought vectors on a wide range of transfer tasks. Much like how computer vision uses ImageNet to obtain features, which can then be transferred to other tasks, our work tends to indicate the suitability of natural language inference for transfer learning to other NLP tasks. Our encoder is publicly available.

The dominant approach for many NLP tasks are recurrent neural networks, in particular LSTMs, and convolutional neural networks. However, these architectures are rather shallow in comparison to the deep convolutional networks which have pushed the state-of-the-art in computer vision. We present a new architecture (VDCNN) for text processing which operates directly at the character level and uses only small convolutions and pooling operations. We are able to show that the performance of this model increases with the depth: using up to 29 convolutional layers, we report improvements over the state-of-the-art on several public text classification tasks. To the best of our knowledge, this is the first time that very deep convolutional nets have been applied to text processing.

In this paper, we use the framework of neural machine translation to learn joint sentence representations across six very different languages. Our aim is that a representation which is independent of the language, is likely to capture the underlying semantics. We define a new cross-lingual similarity measure, compare up to 1.4M sentence representations and study the characteristics of close sentences. We provide experimental evidence that sentences that are close in embedding space are indeed semantically highly related, but often have quite different structure and syntax. These relations also hold when comparing sentences in different languages.

This paper describes the Spoken Language Translation system developed by the LIUM for the IWSLT 2014 evaluation campaign. We participated in two of the proposed tasks: (i) the Automatic Speech Recognition task (ASR) in two languages, Italian with the Vecsys company, and English alone, (ii) the English to French Spoken Language Translation task (SLT). We present the approaches and specificities found in our systems, as well as the results from the evaluation campaign.

This paper describes the development of a statistical machine translation system between French and English for scientific papers. This system will be closely integrated into the French HAL open archive, a collection of more than 100.000 scientific papers. We describe the creation of in-domain parallel and monolingual corpora, the development of a domain specific translation system with the created resources, and its adaptation using monolingual resources only. These techniques allowed us to improve a generic system by more than 10 BLEU points.

The standard procedure to train the translation model of a phrase-based SMT system is to concatenate all available parallel data, to perform word alignment, to extract phrase pairs and to calculate translation probabilities by simple relative frequency. However, parallel data is quite inhomogeneous in many practical applications with respect to several factors like data source, alignment quality, appropriateness to the task, etc. We propose a general framework to take into account these factors during the calculation of the phrase-table, e.g. by better distributing the probability mass of the individual phrase pairs. No additional feature functions are needed. We report results on two well-known tasks: the IWSLT’11 and WMT’11 evaluations, in both conditions translating from English to French. We give detailed results for different functions to weight the bitexts. Our best systems improve a strong baseline by up to one BLEU point without any impact on the computational complexity during training or decoding.

Transliteration is the process of writing a word (mainly proper noun) from one language in the alphabet of another language. This process requires mapping the pronunciation of the word from the source language to the closest possible pronunciation in the target language. In this paper we introduce a new semi-supervised transliteration mining method for parallel and comparable corpora. The method is mainly based on a new suggested Three Levels of Similarity (TLS) scores to extract the transliteration pairs. The first level calculates the similarity of of all vowel letters and consonants letters. The second level calculates the similarity of long vowels and vowel letters at beginning and end position of the words and consonants letters. The third level calculates the similarity consonants letters only. We applied our method on Arabic-English parallel and comparable corpora. We evaluated the extracted transliteration pairs using a statistical based transliteration system. This system is built using letters instead or words as tokens. The transliteration system achieves an accuracy of 0.50 and a mean F-score 0.8958 when trained on transliteration pairs extracted from a parallel corpus. The accuracy is 0.30 and the mean F-score 0.84 when we used instead a comparable corpus to automatically extract the transliteration pairs. This shows that the proposed semi-supervised transliteration mining algorithm is effective and can be applied to other language pairs. We also evaluated two segmentation techniques and reported the impact on the transliteration performance.

It is well known that statistical machine translation systems perform best when they are adapted to the task. In this paper we propose new methods to quickly perform incremental adaptation without the need to obtain word-by-word alignments from GIZA or similar tools. The main idea is to use an automatic translation as pivot to infer alignments between the source sentence and the reference translation, or user correction. We compared our approach to the standard method to perform incremental re-training. We achieve similar results in the BLEU score using less computational resources. Fast retraining is particularly interesting when we want to almost instantly integrate user feed-back, for instance in a post-editing context or machine translation assisted CAT tool. We also explore several methods to combine the translation models.

This paper describes the three systems developed by the LIUM for the IWSLT 2011 evaluation campaign. We participated in three of the proposed tasks, namely the Automatic Speech Recognition task (ASR), the ASR system combination task (ASR_SC) and the Spoken Language Translation task (SLT), since these tasks are all related to speech translation. We present the approaches and specificities we developed on each task.

Les performances d’un système de traduction statistique dépendent beaucoup de la qualité et de la quantité des données d’apprentissage disponibles. La plupart des textes parallèles librement disponibles proviennent d’organisations internationales. Le jargon observé dans ces textes n’est pas très adapté pour construire un système de traduction pour d’autres domaines. Nous présentons dans cet article une technique pour adapter le modèle de traduction à un domaine différent en utilisant des textes dans la langue source uniquement. Nous obtenons des améliorations significatives du score BLEU dans des systèmes de traduction de l’arabe vers le français et vers l’anglais.

This paper describes the systems developed by the LIUM laboratory for the 2009 IWSLT evaluation. We participated in the Arabic and Chinese to English BTEC tasks. We developed three different systems: a statistical phrase-based system using the Moses toolkit, an Statistical Post-Editing system and a hierarchical phrase-based system based on Joshua. A continuous space language model was deployed to improve the modeling of the target language. These systems are combined by a confusion network based approach.

This paper describes the system developed by the LIUM laboratory for the 2008 IWSLT evaluation. We only participated in the Arabic/English BTEC task. We developed a statistical phrase-based system using the Moses toolkit and SYSTRAN’s rule-based translation system to perform a morphological decomposition of the Arabic words. A continuous space language model was deployed to improve the modeling of the target language. Both approaches achieved significant improvements in the BLEU score. The system achieves a score of 49.4 on the test set of the 2008 IWSLT evaluation.

Sentence-aligned bilingual texts are a crucial resource to build statistical machine translation (SMT) systems. In this paper we propose to apply lightly-supervised training to produce additional parallel data. The idea is to translate large amounts of monolingual data (up to 275M words) with an SMT system, and to use those as additional training data. Results are reported for the translation from French into English. We consider two setups: first the intial SMT system is only trained with a very limited amount of human-produced translations, and then the case where we have more than 100 million words. In both conditions, lightly-supervised training achieves significant improvements of the BLEU score.

La traduction automatique statistique par séquences de mots est une voie prometteuse. Nous présentons dans cet article deux évolutions complémentaires. La première permet une modélisation de la langue cible dans un espace continu. La seconde intègre des catégories morpho-syntaxiques aux unités manipulées par le modèle de traduction. Ces deux approches sont évaluées sur la tâche Tc-Star. Les résultats les plus intéressants sont obtenus par la combinaison de ces deux méthodes.

This paper describes TALPtuples, the 2007 N-gram-based statistical machine translation system developed at the TALP Research Center of the UPC (Universitat Polite`cnica de Catalunya) in Barcelona. Emphasis is put on improvements and extensions of the system of previous years. Mainly, these include optimizing alignment parameters in function of translation metric scores and rescoring with a neural network language model. Results on two translation directions are reported, namely from Arabic and Chinese into English, thoroughly explaining all language-related preprocessing and translation schemes.