This paper presents Z-Code++, a new pre-trained language model optimized for abstractive text summarization. The model extends the state-of-the-art encoder-decoder model using three techniques. First, we use a two-phase pre-training to improve the model’s performance on low-resource summarization tasks. The model is first pre-trained using text corpora for language understanding, then is continually pre-trained on summarization corpora for grounded text generation. Second, we replace self-attention layers in the encoder with disentangled attention layers, where each word is represented using two vectors that encode its content and position, respectively. Third, we use fusion-in-encoder, a simple yet effective method of encoding long sequences in a hierarchical manner. Z-Code++ createsa new state-of-the-art on 9 of 13 text summarization tasks across 5 languages. Our model is parameter-efficient in that it outperforms the 600x larger PaLM540B on XSum, and the finetuned 200x larger GPT3175B on SAMSum. In zero-shot and few-shot settings, our model substantially outperforms the competing models.

Large Language Models (LLMs) such as GPT-3 have emerged as general-purpose language models capable of addressing many natural language generation or understanding tasks. On the task of Machine Translation (MT), multiple works have investigated few-shot prompting mechanisms to elicit better translations from LLMs. However, there has been relatively little investigation on how such translations differ qualitatively from the translations generated by standard Neural Machine Translation (NMT) models. In this work, we investigate these differences in terms of the literalness of translations produced by the two systems. Using literalness measures involving word alignment and monotonicity, we find that translations out of English (E-X) from GPTs tend to be less literal, while exhibiting similar or better scores on MT quality metrics. We demonstrate that this finding is borne out in human evaluations as well. We then show that these differences are especially pronounced when translating sentences that contain idiomatic expressions.

Multilingual Neural Machine Translation (MNMT) enables one system to translate sentences from multiple source languages to multiple target languages, greatly reducing deployment costs compared with conventional bilingual systems. The MNMT training benefit, however, is often limited to many-to-one directions. The model suffers from poor performance in one-to-many and many-to-many with zero-shot setup. To address this issue, this paper discusses how to practically build MNMT systems that serve arbitrary X-Y translation directions while leveraging multilinguality with a two-stage training strategy of pretraining and finetuning. Experimenting with the WMT’21 multilingual translation task, we demonstrate that our systems outperform the conventional baselines of direct bilingual models and pivot translation models for most directions, averagely giving +6.0 and +4.1 BLEU, without the need for architecture change or extra data collection. Moreover, we also examine our proposed approach in an extremely large-scale data setting to accommodate practical deployment scenarios.

Multilingual Neural Machine Translation has been showing great success using transformer models. Deploying these models is challenging because they usually require large vocabulary (vocab) sizes for various languages. This limits the speed of predicting the output tokens in the last vocab projection layer. To alleviate these challenges, this paper proposes a fast vocabulary projection method via clustering which can be used for multilingual transformers on GPUs. First, we offline split the vocab search space into disjoint clusters given the hidden context vector of the decoder output, which results in much smaller vocab columns for vocab projection. Second, at inference time, the proposed method predicts the clusters and candidate active tokens for hidden context vectors at the vocab projection. This paper also includes analysis of different ways of building these clusters in multilingual settings. Our results show end-to-end speed gains in float16 GPU inference up to 25% while maintaining the BLEU score and slightly increasing memory cost. The proposed method speeds up the vocab projection step itself by up to 2.6x. We also conduct an extensive human evaluation to verify the proposed method preserves the quality of the translations from the original model.

This paper proposes a simple and effective method to improve direct translation for the zero-shot case and when direct data is available. We modify the input tokens at both the encoder and decoder to include signals for the source and target languages. We show a performance gain when training from scratch, or finetuning a pretrained model with the proposed setup. In in-house experiments, our method shows nearly a 10.0 BLEU points difference depending on the stoppage criteria. In a WMT-based setting, we see 1.3 and 0.4 BLEU points improvement for the zero-shot setting, and when using direct data for training, respectively, while from-English performance improves by 4.17 and 0.85 BLEU points. In the low-resource setting, we see a 1.5 ∼ 1.7 point improvement when finetuning on directly translated domain data.

Mixture of Experts (MoE) models with conditional execution of sparsely activated layers has enabled training models with a much larger number of parameters. As a result, these models have achieved significantly better quality on various natural language processing tasks including machine translation. However, it remains challenging to deploy such models in real-life scenarios due to the large memory requirements and inefficient inference. In this work, we introduce a highly efficient inference framework with several optimization approaches to accelerate the computation of sparse models and cut down the memory consumption significantly. While we achieve up to 26x speed-up in terms of throughput, we also reduce the model size almost to one eighth of the original 32-bit float model by quantizing expert weights into 4-bit integers. As a result, we are able to deploy 136x larger models with 27% less cost and significantly better quality with large scale MoE model deployment compared to the existing solutions. This enables a paradigm shift in deploying large scale multilingual MoE transformers models instead of distilling into dozens of smaller models per language or task.

This paper describes the Microsoft Egypt Development Center (EgDC) submission to the constrained track of WMT21 shared news translation task. We focus on the three relatively low resource language pairs Bengali ↔ Hindi, English ↔ Hausa and Xhosa ↔ Zulu. To overcome the limitation of relatively low parallel data we train a multilingual model using a multitask objective employing both parallel and monolingual data. In addition, we augment the data using back translation. We also train a bilingual model incorporating back translation and knowledge distillation then combine the two models using sequence-to-sequence mapping. We see around 70% relative gain in BLEU point for En ↔ Ha and around 25% relative improvements for Bn ↔ Hi and Xh ↔ Zu compared to bilingual baselines.

This report describes Microsoft’s machine translation systems for the WMT21 shared task on large-scale multilingual machine translation. We participated in all three evaluation tracks including Large Track and two Small Tracks where the former one is unconstrained and the latter two are fully constrained. Our model submissions to the shared task were initialized with DeltaLM, a generic pre-trained multilingual encoder-decoder model, and fine-tuned correspondingly with the vast collected parallel data and allowed data sources according to track settings, together with applying progressive learning and iterative back-translation approaches to further improve the performance. Our final submissions ranked first on three tracks in terms of the automatic evaluation metric.

Multilingual pre-trained models have demonstrated their effectiveness in many multilingual NLP tasks and enabled zero-shot or few-shot transfer from high-resource languages to low-resource ones. However, due to significant typological differences and contradictions between some languages, such models usually perform poorly on many languages and cross-lingual settings, which shows the difficulty of learning a single model to handle massive diverse languages well at the same time. To alleviate this issue, we present a new multilingual pre-training pipeline. We propose to generate language representation from multilingual pre-trained model and conduct linguistic analysis to show that language representation similarity reflects linguistic similarity from multiple perspectives, including language family, geographical sprachbund, lexicostatistics, and syntax. Then we cluster all the target languages into multiple groups and name each group as a representation sprachbund. Thus, languages in the same representation sprachbund are supposed to boost each other in both pre-training and fine-tuning as they share rich linguistic similarity. We pre-train one multilingual model for each representation sprachbund. Experiments are conducted on cross-lingual benchmarks and significant improvements are achieved compared to strong baselines.

Multilingual Neural Machine Translation (NMT) enables one model to serve all translation directions, including ones that are unseen during training, i.e. zero-shot translation. Despite being theoretically attractive, current models often produce low quality translations – commonly failing to even produce outputs in the right target language. In this work, we observe that off-target translation is dominant even in strong multilingual systems, trained on massive multilingual corpora. To address this issue, we propose a joint approach to regularize NMT models at both representation-level and gradient-level. At the representation level, we leverage an auxiliary target language prediction task to regularize decoder outputs to retain information about the target language. At the gradient level, we leverage a small amount of direct data (in thousands of sentence pairs) to regularize model gradients. Our results demonstrate that our approach is highly effective in both reducing off-target translation occurrences and improving zero-shot translation performance by +5.59 and +10.38 BLEU on WMT and OPUS datasets respectively. Moreover, experiments show that our method also works well when the small amount of direct data is not available.

Transformer-based models are the state-of-the-art for Natural Language Understanding (NLU) applications. Models are getting bigger and better on various tasks. However, Transformer models remain computationally challenging since they are not efficient at inference-time compared to traditional approaches. In this paper, we present FastFormers, a set of recipes to achieve efficient inference-time performance for Transformer-based models on various NLU tasks. We show how carefully utilizing knowledge distillation, structured pruning and numerical optimization can lead to drastic improvements on inference efficiency. We provide effective recipes that can guide practitioners to choose the best settings for various NLU tasks and pretrained models. Applying the proposed recipes to the SuperGLUE benchmark, we achieve from 9.8x up to 233.9x speed-up compared to out-of-the-box models on CPU. On GPU, we also achieve up to 12.4x speed-up with the presented methods. We show that FastFormers can drastically reduce cost of serving 100 million requests from 4,223 USD to just 18 USD on an Azure F16s_v2 instance. This translates to a sustainable runtime by reducing energy consumption 6.9x - 125.8x according to the metrics used in the SustaiNLP 2020 shared task.

We present META-MT, a meta-learning approach to adapt Neural Machine Translation (NMT) systems in a few-shot setting. META-MT provides a new approach to make NMT models easily adaptable to many target do- mains with the minimal amount of in-domain data. We frame the adaptation of NMT systems as a meta-learning problem, where we learn to adapt to new unseen domains based on simulated offline meta-training domain adaptation tasks. We evaluate the proposed meta-learning strategy on ten domains with general large scale NMT systems. We show that META-MT significantly outperforms classical domain adaptation when very few in- domain examples are available. Our experiments shows that META-MT can outperform classical fine-tuning by up to 2.5 BLEU points after seeing only 4, 000 translated words (300 parallel sentences).

This paper presents the description of our submission to WMT20 sentence filtering task. We combine scores from custom LASER built for each source language, a classifier built to distinguish positive and negative pairs and the original scores provided with the task. For the mBART setup, provided by the organizers, our method shows 7% and 5% relative improvement, over the baseline, in sacreBLEU score on the test set for Pashto and Khmer respectively.

While monolingual data has been shown to be useful in improving bilingual neural machine translation (NMT), effectively and efficiently leveraging monolingual data for Multilingual NMT (MNMT) systems is a less explored area. In this work, we propose a multi-task learning (MTL) framework that jointly trains the model with the translation task on bitext data and two denoising tasks on the monolingual data. We conduct extensive empirical studies on MNMT systems with 10 language pairs from WMT datasets. We show that the proposed approach can effectively improve the translation quality for both high-resource and low-resource languages with large margin, achieving significantly better results than the individual bilingual models. We also demonstrate the efficacy of the proposed approach in the zero-shot setup for language pairs without bitext training data. Furthermore, we show the effectiveness of MTL over pre-training approaches for both NMT and cross-lingual transfer learning NLU tasks; the proposed approach outperforms massive scale models trained on single task.

Modern NLP applications have enjoyed a great boost utilizing neural networks models. Such deep neural models, however, are not applicable to most human languages due to the lack of annotated training data for various NLP tasks. Cross-lingual transfer learning (CLTL) is a viable method for building NLP models for a low-resource target language by leveraging labeled data from other (source) languages. In this work, we focus on the multilingual transfer setting where training data in multiple source languages is leveraged to further boost target language performance. Unlike most existing methods that rely only on language-invariant features for CLTL, our approach coherently utilizes both language-invariant and language-specific features at instance level. Our model leverages adversarial networks to learn language-invariant features, and mixture-of-experts models to dynamically exploit the similarity between the target language and each individual source language. This enables our model to learn effectively what to share between various languages in the multilingual setup. Moreover, when coupled with unsupervised multilingual embeddings, our model can operate in a zero-resource setting where neither target language training data nor cross-lingual resources are available. Our model achieves significant performance gains over prior art, as shown in an extensive set of experiments over multiple text classification and sequence tagging tasks including a large-scale industry dataset.

This paper describes the submissions of the “Marian” team to the WNGT 2019 efficiency shared task. Taking our dominating submissions to the previous edition of the shared task as a starting point, we develop improved teacher-student training via multi-agent dual-learning and noisy backward-forward translation for Transformer-based student models. For efficient CPU-based decoding, we propose pre-packed 8-bit matrix products, improved batched decoding, cache-friendly student architectures with parameter sharing and light-weight RNN-based decoder architectures. GPU-based decoding benefits from the same architecture changes, from pervasive 16-bit inference and concurrent streams. These modifications together with profiler-based C++ code optimization allow us to push the Pareto frontier established during the 2018 edition towards 24x (CPU) and 14x (GPU) faster models at comparable or higher BLEU values. Our fastest CPU model is more than 4x faster than last year’s fastest submission at more than 3 points higher BLEU. Our fastest GPU model at 1.5 seconds translation time is slightly faster than last year’s fastest RNN-based submissions, but outperforms them by more than 4 BLEU and 10 BLEU points respectively.

In this paper, we report our system submissions to all 6 tracks of the WNGT 2019 shared task on Document-Level Generation and Translation. The objective is to generate a textual document from either structured data: generation task, or a document in a different language: translation task. For the translation task, we focused on adapting a large scale system trained on WMT data by fine tuning it on the RotoWire data. For the generation task, we participated with two systems based on a selection and planning model followed by (a) a simple language model generation, and (b) a GPT-2 pre-trained language model approach. The selection and planning module chooses a subset of table records in order, and the language models produce text given such a subset.

Parallel corpora available for building machine translation (MT) models for dialectal Arabic (DA) are rather limited. The scarcity of resources has prompted the use of Modern Standard Arabic (MSA) abundant resources to complement the limited dialectal resource. However, dialectal clitics often differ between MSA and DA. This paper compares morphology-aware DA word segmentation to other word segmentation approaches like Byte Pair Encoding (BPE) and Sub-word Regularization (SR). A set of experiments conducted on Egyptian Arabic (EA), Levantine Arabic (LA), and Gulf Arabic (GA) show that a sufficiently accurate morphology-aware segmentation used in conjunction with BPE outperforms the other word segmentation approaches.

In this paper, we propose a new universal machine translation approach focusing on languages with a limited amount of parallel data. Our proposed approach utilizes a transfer-learning approach to share lexical and sentence level representations across multiple source languages into one target language. The lexical part is shared through a Universal Lexical Representation to support multi-lingual word-level sharing. The sentence-level sharing is represented by a model of experts from all source languages that share the source encoders with all other languages. This enables the low-resource language to utilize the lexical and sentence representations of the higher resource languages. Our approach is able to achieve 23 BLEU on Romanian-English WMT2016 using a tiny parallel corpus of 6k sentences, compared to the 18 BLEU of strong baseline system which uses multi-lingual training and back-translation. Furthermore, we show that the proposed approach can achieve almost 20 BLEU on the same dataset through fine-tuning a pre-trained multi-lingual system in a zero-shot setting.

In this paper, we introduce a novel approach to generate synthetic data for training Neural Machine Translation systems. The proposed approach supports language variants and dialects with very limited parallel training data. This is achieved using a seed data to project words from a closely-related resource-rich language to an under-resourced language variant via word embedding representations. The proposed approach is based on localized embedding projection of distributed representations which utilizes monolingual embeddings and approximate nearest neighbors queries to transform parallel data across language variants. Our approach is language independent and can be used to generate data for any variant of the source language such as slang or spoken dialect or even for a different language that is related to the source language. We report experimental results on Levantine to English translation using Neural Machine Translation. We show that the synthetic data can provide significant improvements over a very large scale system by more than 2.8 Bleu points and it can be used to provide a reliable translation system for a spoken dialect which does not have sufficient parallel data.

This paper describes the systems used for the MSR+FBK submission for the SLT track of IWSLT 2013. Starting from a baseline system we made a series of iterative and additive improvements, including a novel method for processing bilingual data used to train MT systems for use on ASR output. Our primary submission is a system combination of five individual systems, combining the output of multiple ASR engines with multiple MT techniques. There are two contrastive submissions to help place the combined system in context. We describe the systems used and present results on the test sets.

In this paper, we give a description of the machine translation (MT) system developed at DCU that was used for our third participation in the evaluation campaign of the International Workshop on Spoken Language Translation (IWSLT 2008). In this participation, we focus on various techniques for word and phrase alignment to improve system quality. Specifically, we try out our word packing and syntax-enhanced word alignment techniques for the Chinese–English task and for the English–Chinese task for the first time. For all translation tasks except Arabic–English, we exploit linguistically motivated bilingual phrase pairs extracted from parallel treebanks. We smooth our translation tables with out-of-domain word translations for the Arabic–English and Chinese–English tasks in order to solve the problem of the high number of out of vocabulary items. We also carried out experiments combining both in-domain and out-of-domain data to improve system performance and, finally, we deploy a majority voting procedure combining a language model-based method and a translation-based method for case and punctuation restoration. We participated in all the translation tasks and translated both the single-best ASR hypotheses and the correct recognition results. The translation results confirm that our new word and phrase alignment techniques are often helpful in improving translation quality, and the data combination method we proposed can significantly improve system performance.

In this paper, we give a description of the machine translation system developed at DCU that was used for our second participation in the evaluation campaign of the International Workshop on Spoken Language Translation (IWSLT 2007). In this participation, we focus on some new methods to improve system quality. Specifically, we try our word packing technique for different language pairs, we smooth our translation tables with out-of-domain word translations for the Arabic–English and Chinese–English tasks in order to solve the high number of out of vocabulary items, and finally we deploy a translation-based model for case and punctuation restoration. We participated in both the classical and challenge tasks for the following translation directions: Chinese–English, Japanese–English and Arabic–English. For the last two tasks, we translated both the single-best ASR hypotheses and the correct recognition results; for Chinese–English, we just translated the correct recognition results. We report the results of the system for the provided evaluation sets, together with some additional experiments carried out following identification of some simple tokenisation errors in the official runs.