Arul Menezes

Shufang Xie, Huishuai Zhang, Junliang Guo et al. · microsoft-research

Transformer networks have become the preferred architecture for many tasks due to their state-of-the-art performance. However, the optimal way to implement residual connections in Transformer, which are essential for effective training, is still debated. Two widely used variants are the Post-Layer-Normalization (Post-LN) and Pre-Layer-Normalization (Pre-LN) Transformers, which apply layer normalization after each residual block's output or before each residual block's input, respectively. While both variants enjoy their advantages, they also suffer from severe limitations: Post-LN causes gradient vanishing issue that hinders training deep Transformers, and Pre-LN causes representation collapse issue that limits model capacity. In this paper, we propose ResiDual, a novel Transformer architecture with Pre-Post-LN (PPLN), which fuses the connections in Post-LN and Pre-LN together and inherits their advantages while avoids their limitations. We conduct both theoretical analyses and empirical experiments to verify the effectiveness of ResiDual. Theoretically, we prove that ResiDual has a lower bound on the gradient to avoid the vanishing issue due to the residual connection from Pre-LN. Moreover, ResiDual also has diverse model representations to avoid the collapse issue due to the residual connection from Post-LN. Empirically, ResiDual outperforms both Post-LN and Pre-LN on several machine translation benchmarks across different network depths and data sizes. Thanks to the good theoretical and empirical performance, ResiDual Transformer can serve as a foundation architecture for different AI models (e.g., large language models). Our code is available at https://github.com/microsoft/ResiDual.

Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao et al. · allen-ai, amazon-science

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Vikas Raunak, Arul Menezes · microsoft-research

Memorization presents a challenge for several constrained Natural Language Generation (NLG) tasks such as Neural Machine Translation (NMT), wherein the proclivity of neural models to memorize noisy and atypical samples reacts adversely with the noisy (web crawled) datasets. However, previous studies of memorization in constrained NLG tasks have only focused on counterfactual memorization, linking it to the problem of hallucinations. In this work, we propose a new, inexpensive algorithm for extractive memorization (exact training data generation under insufficient context) in constrained sequence generation tasks and use it to study extractive memorization and its effects in NMT. We demonstrate that extractive memorization poses a serious threat to NMT reliability by qualitatively and quantitatively characterizing the memorized samples as well as the model behavior in their vicinity. Based on empirical observations, we develop a simple algorithm which elicits non-memorized translations of memorized samples from the same model, for a large fraction of such samples. Finally, we show that the proposed algorithm could also be leveraged to mitigate memorization in the model through finetuning. We have released the code to reproduce our results at https://github.com/vyraun/Finding-Memo.

Yihan Wu, Junliang Guo, Xu Tan et al. · microsoft-research

Video dubbing aims to translate the original speech in a film or television program into the speech in a target language, which can be achieved with a cascaded system consisting of speech recognition, machine translation and speech synthesis. To ensure the translated speech to be well aligned with the corresponding video, the length/duration of the translated speech should be as close as possible to that of the original speech, which requires strict length control. Previous works usually control the number of words or characters generated by the machine translation model to be similar to the source sentence, without considering the isochronicity of speech as the speech duration of words/characters in different languages varies. In this paper, we propose a machine translation system tailored for the task of video dubbing, which directly considers the speech duration of each token in translation, to match the length of source and target speech. Specifically, we control the speech length of generated sentence by guiding the prediction of each word with the duration information, including the speech duration of itself as well as how much duration is left for the remaining words. We design experiments on four language directions (German -> English, Spanish -> English, Chinese <-> English), and the results show that the proposed method achieves better length control ability on the generated speech than baseline methods. To make up the lack of real-world datasets, we also construct a real-world test set collected from films to provide comprehensive evaluations on the video dubbing task.

Vikas Raunak, Matt Post, Arul Menezes · microsoft-research

Traditional machine translation (MT) metrics provide an average measure of translation quality that is insensitive to the long tail of behavioral problems in MT. Examples include translation of numbers, physical units, dropped content and hallucinations. These errors, which occur rarely and unpredictably in Neural Machine Translation (NMT), greatly undermine the reliability of state-of-the-art MT systems. Consequently, it is important to have visibility into these problems during model development. Towards this direction, we introduce SALTED, a specifications-based framework for behavioral testing of MT models that provides fine-grained views of salient long-tail errors, permitting trustworthy visibility into previously invisible problems. At the core of our approach is the development of high-precision detectors that flag errors (or alternatively, verify output correctness) between a source sentence and a system output. We demonstrate that such detectors could be used not just to identify salient long-tail errors in MT systems, but also for higher-recall filtering of the training data, fixing targeted errors with model fine-tuning in NMT and generating novel data for metamorphic testing to elicit further bugs in models.

Vikas Raunak, Arul Menezes · microsoft-research

Large sequence to sequence models for tasks such as Neural Machine Translation (NMT) are usually trained over hundreds of millions of samples. However, training is just the origin of a model's life-cycle. Real-world deployments of models require further behavioral adaptations as new requirements emerge or shortcomings become known. Typically, in the space of model behaviors, behavior deletion requests are addressed through model retrainings whereas model finetuning is done to address behavior addition requests, both procedures being instances of data-based model intervention. In this work, we present a preliminary study investigating rank-one editing as a direct intervention method for behavior deletion requests in encoder-decoder transformer models. We propose four editing tasks for NMT and show that the proposed editing algorithm achieves high efficacy, while requiring only a single instance of positive example to fix an erroneous (negative) model behavior.

Vikas Raunak, Matt Post, Arul Menezes · microsoft-research

In this work, we present some recommendations on the evaluation of state-of-the-art generative models for constrained generation tasks. The progress on generative models has been rapid in recent years. These large-scale models have had three impacts: firstly, the fluency of generation in both language and vision modalities has rendered common average-case evaluation metrics much less useful in diagnosing system errors. Secondly, the same substrate models now form the basis of a number of applications, driven both by the utility of their representations as well as phenomena such as in-context learning, which raise the abstraction level of interacting with such models. Thirdly, the user expectations around these models and their feted public releases have made the technical challenge of out of domain generalization much less excusable in practice. Subsequently, our evaluation methodologies haven't adapted to these changes. More concretely, while the associated utility and methods of interacting with generative models have expanded, a similar expansion has not been observed in their evaluation practices. In this paper, we argue that the scale of generative models could be exploited to raise the abstraction level at which evaluation itself is conducted and provide recommendations for the same. Our recommendations are based on leveraging specifications as a powerful instrument to evaluate generation quality and are readily applicable to a variety of tasks.

Vikas Raunak, Hany Hassan Awadalla, Arul Menezes

Most of the recent work in leveraging Large Language Models (LLMs) such as GPT-3 for Machine Translation (MT) has focused on selecting the few-shot samples for prompting. In this work, we try to better understand the role of demonstration attributes for the in-context learning of translations through perturbations of high-quality, in-domain demonstrations. We find that asymmetric perturbation of the source-target mappings yield vastly different results. We show that the perturbation of the source side has surprisingly little impact, while target perturbation can drastically reduce translation quality, suggesting that it is the output text distribution that provides the most important learning signal during in-context learning of translations. We propose a method named Zero-Shot-Context to add this signal automatically in Zero-Shot prompting. We demonstrate that it improves upon the zero-shot translation performance of GPT-3, even making it competitive with few-shot prompted translations.

Guduru Manoj, Neel Prabhanjan Rachamalla, Ashish Kulkarni et al.

In the context of pretraining of Large Language Models (LLMs), synthetic data has emerged as an alternative for generating high-quality pretraining data at scale. This is particularly beneficial in low-resource language settings where the benefits of recent LLMs have been unevenly distributed across languages. In this work, we present a systematic study on the generation and evaluation of synthetic multilingual pretraining data for Indic languages, where we construct a large-scale synthetic dataset BhashaKritika, comprising 540B tokens using 5 different techniques for 10 languages. We explore the impact of grounding generation in documents, personas, and topics. We analyze how language choice, both in the prompt instructions and document grounding, affects data quality, and we compare translations of English content with native generation in Indic languages. To support scalable and language-sensitive evaluation, we introduce a modular quality evaluation pipeline that integrates script and language detection, metadata consistency checks, n-gram repetition analysis, and perplexity-based filtering using KenLM models. Our framework enables robust quality control across diverse scripts and linguistic contexts. Empirical results through model runs reveal key trade-offs in generation strategies and highlight best practices for constructing effective multilingual corpora.

Souvik Rana, Arul Menezes, Ashish Kulkarni et al.

Tokenizers play a crucial role in determining the performance, training efficiency, and the inference cost of Large Language Models (LLMs). Designing effective tokenizers for multilingual LLMs is particularly challenging due to diverse scripts and rich morphological variation. While subword methods such as Byte Pair Encoding (BPE) are widely adopted, their effectiveness in multilingual settings remains underexplored. We present IndicSuperTokenizer, a tokenizer for Indic multilingual LLMs, that combines both subword and multi-word tokenization, along with language-specific pre-tokenization, leading to more linguistically aligned tokens and achieving a new state-of-the-art in fertility score. Evaluated across English, 22 Indian languages and code data, our tokenizer improves the average fertility score by 39.5% over LLaMA4 and by 18% over Sutra (the current best). This translates to 44% improvement in inference throughput over LLaMA4 while maintaining comparable performance on English and Indic benchmarks. We also present detailed ablations across tokenizer training data size, vocabulary size, merging techniques, and pre-tokenization strategies, demonstrating the robustness of our design choices.

Bei Li, Rui Wang, Junliang Guo et al.

Large language models (LLMs) have shown remarkable success across a wide range of natural language generation tasks, where proper prompt designs make great impacts. While existing prompting methods are normally restricted to providing correct information, in this paper, we encourage the model to deliberate by proposing a novel Deliberate then Generate (DTG) prompting framework, which consists of error detection instructions and candidates that may contain errors. DTG is a simple yet effective technique that can be applied to various text generation tasks with minimal modifications. We conduct extensive experiments on 20+ datasets across 7 text generation tasks, including summarization, translation, dialogue, and more. We show that DTG consistently outperforms existing prompting methods and achieves state-of-the-art performance on multiple text generation tasks. We also provide in-depth analyses to reveal the underlying mechanisms of DTG, which may inspire future research on prompting for LLMs.

Vikas Raunak, Arul Menezes, Matt Post et al.

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.

Vikas Raunak, Amr Sharaf, Yiren Wang et al.

While Neural Machine Translation (NMT) represents the leading approach to Machine Translation (MT), the outputs of NMT models still require translation post-editing to rectify errors and enhance quality under critical settings. In this work, we formalize the task of direct translation post-editing with Large Language Models (LLMs) and explore the use of GPT-4 to automatically post-edit NMT outputs across several language pairs. Our results demonstrate that GPT-4 is adept at translation post-editing, producing meaningful and trustworthy edits to translations that help improve its general quality as well as remove different classes of major errors in translations. In particular, human evaluations on assessing edit trustworthiness show that GPT-4 exhibits a large improvement over the prior state-of-the-art LLM. Notably, we improve upon state-of-the-art performance on WMT-22 English-Chinese, English-German, Chinese-English and German-English language pairs using GPT-4 based post-editing, as evaluated by state-of-the-art MT quality metrics. However, we also show that GPT-4 could produce hallucinated edits, thereby urging caution in its use as an expert translation post-editor.

Tom Kocmi, Christian Federmann, Roman Grundkiewicz et al.

Automatic metrics are commonly used as the exclusive tool for declaring the superiority of one machine translation system's quality over another. The community choice of automatic metric guides research directions and industrial developments by deciding which models are deemed better. Evaluating metrics correlations with sets of human judgements has been limited by the size of these sets. In this paper, we corroborate how reliable metrics are in contrast to human judgements on -- to the best of our knowledge -- the largest collection of judgements reported in the literature. Arguably, pairwise rankings of two systems are the most common evaluation tasks in research or deployment scenarios. Taking human judgement as a gold standard, we investigate which metrics have the highest accuracy in predicting translation quality rankings for such system pairs. Furthermore, we evaluate the performance of various metrics across different language pairs and domains. Lastly, we show that the sole use of BLEU impeded the development of improved models leading to bad deployment decisions. We release the collection of 2.3M sentence-level human judgements for 4380 systems for further analysis and replication of our work.

Vikas Raunak, Arul Menezes, Marcin Junczys-Dowmunt

In this work, we study hallucinations in Neural Machine Translation (NMT), which lie at an extreme end on the spectrum of NMT pathologies. Firstly, we connect the phenomenon of hallucinations under source perturbation to the Long-Tail theory of Feldman (2020), and present an empirically validated hypothesis that explains hallucinations under source perturbation. Secondly, we consider hallucinations under corpus-level noise (without any source perturbation) and demonstrate that two prominent types of natural hallucinations (detached and oscillatory outputs) could be generated and explained through specific corpus-level noise patterns. Finally, we elucidate the phenomenon of hallucination amplification in popular data-generation processes such as Backtranslation and sequence-level Knowledge Distillation.

Shuming Ma, Jian Yang, Haoyang Huang et al.

Multilingual machine translation enables a single model to translate between different languages. Most existing multilingual machine translation systems adopt a randomly initialized Transformer backbone. In this work, inspired by the recent success of language model pre-training, we present XLM-T, which initializes the model with an off-the-shelf pretrained cross-lingual Transformer encoder and fine-tunes it with multilingual parallel data. This simple method achieves significant improvements on a WMT dataset with 10 language pairs and the OPUS-100 corpus with 94 pairs. Surprisingly, the method is also effective even upon the strong baseline with back-translation. Moreover, extensive analysis of XLM-T on unsupervised syntactic parsing, word alignment, and multilingual classification explains its effectiveness for machine translation. The code will be at https://aka.ms/xlm-t.

Hany Hassan, Anthony Aue, Chang Chen et al.

Machine translation has made rapid advances in recent years. Millions of people are using it today in online translation systems and mobile applications in order to communicate across language barriers. The question naturally arises whether such systems can approach or achieve parity with human translations. In this paper, we first address the problem of how to define and accurately measure human parity in translation. We then describe Microsoft's machine translation system and measure the quality of its translations on the widely used WMT 2017 news translation task from Chinese to English. We find that our latest neural machine translation system has reached a new state-of-the-art, and that the translation quality is at human parity when compared to professional human translations. We also find that it significantly exceeds the quality of crowd-sourced non-professional translations.