Training stability is of great importance to Transformers. In this work, we investigate the training dynamics of Transformers by examining the evolution of the attention layers. In particular, we track the attention entropy for each attention head during the course of training, which is a proxy for model sharpness. We identify a common pattern across different architectures and tasks, where low attention entropy is accompanied by high training instability, which can take the form of oscillating loss or divergence. We denote the pathologically low attention entropy, corresponding to highly concentrated attention scores, as $\textit{entropy collapse}$. As a remedy, we propose $\sigma$Reparam, a simple and efficient solution where we reparametrize all linear layers with spectral normalization and an additional learned scalar. We demonstrate that $\sigma$Reparam successfully prevents entropy collapse in the attention layers, promoting more stable training. Additionally, we prove a tight lower bound of the attention entropy, which decreases exponentially fast with the spectral norm of the attention logits, providing additional motivation for our approach. We conduct experiments with $\sigma$Reparam on image classification, image self-supervised learning, machine translation, speech recognition, and language modeling tasks. We show that $\sigma$Reparam provides stability and robustness with respect to the choice of hyperparameters, going so far as enabling training (a) a Vision Transformer {to competitive performance} without warmup, weight decay, layer normalization or adaptive optimizers; (b) deep architectures in machine translation and (c) speech recognition to competitive performance without warmup and adaptive optimizers. Code is available at \url{https://github.com/apple/ml-sigma-reparam}.

It is a well-known fact that current AI-based language technology -- language models, machine translation systems, multilingual dictionaries and corpora -- focuses on the world's 2-3% most widely spoken languages. Recent research efforts have attempted to expand the coverage of AI technology to `under-resourced languages.' The goal of our paper is to bring attention to a phenomenon that we call linguistic bias: multilingual language processing systems often exhibit a hardwired, yet usually involuntary and hidden representational preference towards certain languages. Linguistic bias is manifested in uneven per-language performance even in the case of similar test conditions. We show that biased technology is often the result of research and development methodologies that do not do justice to the complexity of the languages being represented, and that can even become ethically problematic as they disregard valuable aspects of diversity as well as the needs of the language communities themselves. As our attempt at building diversity-aware language resources, we present a new initiative that aims at reducing linguistic bias through both technological design and methodology, based on an eye-level collaboration with local communities.

It is well known that AI-based language technology -- large language models, machine translation systems, multilingual dictionaries, and corpora -- is currently limited to 2 to 3 percent of the world's most widely spoken and/or financially and politically best supported languages. In response, recent research efforts have sought to extend the reach of AI technology to ``underserved languages.'' In this paper, we show that many of these attempts produce flawed solutions that adhere to a hard-wired representational preference for certain languages, which we call techno-linguistic bias. Techno-linguistic bias is distinct from the well-established phenomenon of linguistic bias as it does not concern the languages represented but rather the design of the technologies. As we show through the paper, techno-linguistic bias can result in systems that can only express concepts that are part of the language and culture of dominant powers, unable to correctly represent concepts from other communities. We argue that at the root of this problem lies a systematic tendency of technology developer communities to apply a simplistic understanding of diversity which does not do justice to the more profound differences that languages, and ultimately the communities that speak them, embody. Drawing on the concept of epistemic injustice, we point to the broader sociopolitical consequences of the bias we identify and show how it can lead not only to a disregard for valuable aspects of diversity but also to an under-representation of the needs and diverse worldviews of marginalized language communities.

Automatic subtitling is the task of automatically translating the speech of audiovisual content into short pieces of timed text, i.e. subtitles and their corresponding timestamps. The generated subtitles need to conform to space and time requirements, while being synchronised with the speech and segmented in a way that facilitates comprehension. Given its considerable complexity, the task has so far been addressed through a pipeline of components that separately deal with transcribing, translating, and segmenting text into subtitles, as well as predicting timestamps. In this paper, we propose the first direct ST model for automatic subtitling that generates subtitles in the target language along with their timestamps with a single model. Our experiments on 7 language pairs show that our approach outperforms a cascade system in the same data condition, also being competitive with production tools on both in-domain and newly-released out-domain benchmarks covering new scenarios.

Although Neural Machine Translation (NMT) has made remarkable advances (Vaswani et al., 2017), it still requires large amounts of data to induce correct generalizations that characterize human intelligence (Lake et al., 2017). However, such a vast amount of data to make robust, reliable, and fair predictions is not available for low-resource NMT (Koehn and Knowles, 2017). The generalizability of NMT has been extensively studied in prior research, revealing the volatile behaviour of translation outputs when even a single token in the source sentence is modified (Belinkov and Bisk, 2018; Fadaee and Monz, 2020; Li et al., 2021). For instance, in the sentence "smallpox killed billions of people on this planet" from our IWSLT test set, when replacing the noun "smallpox" with another acute disease like "tuberculosis", the model should ideally generate a correct translation by only modifying the relevant part while keeping the rest of the sentence unchanged. However, in many instances, such a small perturbation adversely affects the translation of the entire sentence, highlighting the limited generalization and robustness of existing NMT models (Fadaee and Monz, 2020). Compositionality is regarded as the most prominent form of generalization that embodies the ability of human intelligence to generalize to new data, tasks, and domains (Schmidhuber, 1990; Lake and Baroni, 2018), while other types mostly focus on the practical considerations across domains, tasks, and languages, model robustness, and structural generalization (Hupkes et al., 2022). Research in compositional generalization has two main aspects: evaluating the current models' compositional abilities as well as improving them.

Language Models today provide a high accuracy across a large number of downstream tasks. However, they remain susceptible to adversarial attacks, particularly against those where the adversarial examples maintain considerable similarity to the original text. Given the multilingual nature of text, the effectiveness of adversarial examples across translations and how machine translations can improve the robustness of adversarial examples remain largely unexplored. In this paper, we present a comprehensive study on the robustness of current text adversarial attacks to round-trip translation. We demonstrate that 6 state-of-the-art text-based adversarial attacks do not maintain their efficacy after round-trip translation. Furthermore, we introduce an intervention-based solution to this problem, by integrating Machine Translation into the process of adversarial example generation and demonstrating increased robustness to round-trip translation. Our results indicate that finding adversarial examples robust to translation can help identify the insufficiency of language models that is common across languages, and motivate further research into multilingual adversarial attacks.

Machine Learning (ML) models are ubiquitous in real-world applications and are a constant focus of research. Modern ML models have become more complex, deeper, and harder to reason about. At the same time, the community has started to realize the importance of protecting the privacy of the training data that goes into these models. Differential Privacy (DP) has become a gold standard for making formal statements about data anonymization. However, while some adoption of DP has happened in industry, attempts to apply DP to real world complex ML models are still few and far between. The adoption of DP is hindered by limited practical guidance of what DP protection entails, what privacy guarantees to aim for, and the difficulty of achieving good privacy-utility-computation trade-offs for ML models. Tricks for tuning and maximizing performance are scattered among papers or stored in the heads of practitioners, particularly with respect to the challenging task of hyperparameter tuning. Furthermore, the literature seems to present conflicting evidence on how and whether to apply architectural adjustments and which components are “safe” to use with DP. In this survey paper, we attempt to create a self-contained guide that gives an in-depth overview of the field of DP ML. We aim to assemble information about achieving the best possible DP ML model with rigorous privacy guarantees. Our target audience is both researchers and practitioners. Researchers interested in DP for ML will benefit from a clear overview of current advances and areas for improvement. We also include theory-focused sections that highlight important topics such as privacy accounting and convergence. For a practitioner, this survey provides a background in DP theory and a clear step-by-step guide for choosing an appropriate privacy definition and approach, implementing DP training, potentially updating the model architecture, and tuning hyperparameters. For both researchers and practitioners, consistently and fully reporting privacy guarantees is critical, so we propose a set of specific best practices for stating guarantees. With sufficient computation and a sufficiently large training set or supplemental nonprivate data, both good accuracy (that is, almost as good as a non-private model) and good privacy can often be achievable. And even when computation and dataset size are limited, there are advantages to training with even a weak (but still finite) formal DP guarantee. Hence, we hope this work will facilitate more widespread deployments of DP ML models.

We present a methodology for gathering a parallel corpus through crowdsourcing, which is more cost-effective than hiring professional translators, albeit at the expense of quality. Additionally, we have made available experimental parallel data collected for Chechen-Russian and Fula-English language pairs.

This paper studies the error metric selection for long-term memory learning in sequence modelling. We examine the bias towards short-term memory in commonly used errors, including mean absolute/squared error. Our findings show that all temporally positive-weighted errors are biased towards short-term memory in learning linear functionals. To reduce this bias and improve long-term memory learning, we propose the use of a temporally rescaled error. In addition to reducing the bias towards short-term memory, this approach can also alleviate the vanishing gradient issue. We conduct numerical experiments on different long-memory tasks and sequence models to validate our claims. Numerical results confirm the importance of appropriate temporally rescaled error for effective long-term memory learning. To the best of our knowledge, this is the first work that quantitatively analyzes different errors' memory bias towards short-term memory in sequence modelling.

Modeling discourse -- the linguistic phenomena that go beyond individual sentences, is a fundamental yet challenging aspect of natural language processing (NLP). However, existing evaluation benchmarks primarily focus on the evaluation of inter-sentence properties and overlook critical discourse phenomena that cross sentences. To bridge the gap, we propose Disco-Bench, a benchmark that can evaluate intra-sentence discourse properties across a diverse set of NLP tasks, covering understanding, translation, and generation. Disco-Bench consists of 9 document-level testsets in the literature domain, which contain rich discourse phenomena (e.g. cohesion and coherence) in Chinese and/or English. For linguistic analysis, we also design a diagnostic test suite that can examine whether the target models learn discourse knowledge. We totally evaluate 20 general-, in-domain and commercial models based on Transformer, advanced pretraining architectures and large language models (LLMs). Our results show (1) the challenge and necessity of our evaluation benchmark; (2) fine-grained pretraining based on literary document-level training data consistently improves the modeling of discourse information. We will release the datasets, pretrained models, and leaderboard, which we hope can significantly facilitate research in this field: https://github.com/longyuewangdcu/Disco-Bench.