Ensembles of machine learning models yield improved system performance as well as robust and interpretable uncertainty estimates; however, their inference costs may often be prohibitively high. Ensemble Distribution Distillation is an approach that allows a single model to efficiently capture both the predictive performance and uncertainty estimates of an ensemble. For classification, this is achieved by training a Dirichlet distribution over the ensemble members' output distributions via the maximum likelihood criterion. Although theoretically principled, this criterion exhibits poor convergence when applied to large-scale tasks where the number of classes is very high. In our work, we analyze this effect and show that for the Dirichlet log-likelihood criterion classes with low probability induce larger gradients than high-probability classes. This forces the model to focus on the distribution of the ensemble tail-class probabilities. We propose a new training objective which minimizes the reverse KL-divergence to a Proxy-Dirichlet target derived from the ensemble. This loss resolves the gradient issues of Ensemble Distribution Distillation, as we demonstrate both theoretically and empirically on the ImageNet and WMT17 En-De datasets containing 1000 and 40,000 classes, respectively.

In the past 10 years, hundreds of projects have applied artificial intelligence (AI) to creating social good. The right tool applied to an appropriate problem has the potential to drastically improve millions of lives through better service delivery and better-informed policy design. But what kind of investments do AI solutions need to be successful, and which applications have the most potential for social impact? AI excels at helping humans harness large-scale or complex data to predict, categorize, or optimize at a scale and speed beyond human ability. We believe that more targeted, sustained investments in AI for social impact (sometimes called "AI for good")--rather than multiple, short-term grants across a variety of areas--are important for two reasons.

Questions of fairness, robustness, and transparency are paramount to address before deploying NLP systems. Central to these concerns is the question of reliability: Can NLP systems reliably treat different demographics fairly and function correctly in diverse and noisy environments? To address this, we argue for the need for reliability testing and contextualize it among existing work on improving accountability. We show how adversarial attacks can be reframed for this goal, via a framework for developing reliability tests. We argue that Figure 1: How DOCTOR can integrate with existing reliability testing -- with an emphasis on interdisciplinary system development workflows. Test (left) and system collaboration -- will enable rigorous development (right) take place in parallel, separate and targeted testing, and aid in the enactment teams. Reliability tests can thus be constructed independent and enforcement of industry standards. of the system development team, either by an internal "red team" or by independent auditors.

Generating accurate terminology is a crucial component for the practicality and reliability of neural machine translation (NMT) systems. To address this, lexically constrained NMT explores various methods to ensure pre-specified words and phrases to appear in the translations. In many cases, however, those methods are evaluated on general domain corpora, where the terms are mostly uni- and bi-grams (>98%). In this paper, we instead tackle a more challenging setup consisting of domain-specific corpora with much longer n-gram and highly specialized terms. To encourage span-level representations in generation, we additionally impose a source-sentence conditioned masked span prediction loss in the decoder and observe improvements on both terminology translation as well as BLEU scores. Experimental results on three domain-specific corpora in two language pairs demonstrate that the proposed training scheme can improve the performance of existing lexically constrained methods that can operate both with or without a term dictionary at test time.

Signed languages are the primary means of communication for many deaf and hard of hearing individuals. Since signed languages exhibit all the fundamental linguistic properties of natural language, we believe that tools and theories of Natural Language Processing (NLP) are crucial towards its modeling. However, existing research in Sign Language Processing (SLP) seldom attempt to explore and leverage the linguistic organization of signed languages. This position paper calls on the NLP community to include signed languages as a research area with high social and scientific impact. We first discuss the linguistic properties of signed languages to consider during their modeling. Then, we review the limitations of current SLP models and identify the open challenges to extend NLP to signed languages. Finally, we urge (1) the adoption of an efficient tokenization method; (2) the development of linguistically-informed models; (3) the collection of real-world signed language data; (4) the inclusion of local signed language communities as an active and leading voice in the direction of research.

In the field of machine learning, the well-trained model is assumed to be able to recover the training labels, i.e. the synthetic labels predicted by the model should be as close to the ground-truth labels as possible. Inspired by this, we propose a self-guided curriculum strategy to encourage the learning of neural machine translation (NMT) models to follow the above recovery criterion, where we cast the recovery degree of each training example as its learning difficulty. Specifically, we adopt the sentence level BLEU score as the proxy of recovery degree. Different from existing curricula relying on linguistic prior knowledge or third-party language models, our chosen learning difficulty is more suitable to measure the degree of knowledge mastery of the NMT models. Experiments on translation benchmarks, including WMT14 English$\Rightarrow$German and WMT17 Chinese$\Rightarrow$English, demonstrate that our approach can consistently improve translation performance against strong baseline Transformer.

The data scarcity in low-resource languages has become a bottleneck to building robust neural machine translation systems. Fine-tuning a multilingual pre-trained model (e.g., mBART (Liu et al., 2020)) on the translation task is a good approach for low-resource languages; however, its performance will be greatly limited when there are unseen languages in the translation pairs. In this paper, we present a continual pre-training (CPT) framework on mBART to effectively adapt it to unseen languages. We first construct noisy mixed-language text from the monolingual corpus of the target language in the translation pair to cover both the source and target languages, and then, we continue pre-training mBART to reconstruct the original monolingual text. Results show that our method can consistently improve the fine-tuning performance upon the mBART baseline, as well as other strong baselines, across all tested low-resource translation pairs containing unseen languages. Furthermore, our approach also boosts the performance on translation pairs where both languages are seen in the original mBART's pre-training. The code is available at https://github.com/zliucr/cpt-nmt.

Language is perhaps the most defining factor of humankind. What makes humans different from other animals on the planet is our ability to speak out and communicate via framed words and sentences. The language of a population is one of the most defining factors across countries and nationalities, regions, and cultures. It can define the history, sociocultural situation, and even geographic diversity. From ancient times, there has been a trend for people to understand the language of one another. History traces back to Greeks and Romans traveling all across the world to discover, decipher and translate languages to find out the cultural, political, and social situations from one era to another.

Understanding voluminous historical records provides clues on the past in various aspects, such as social and political issues and even natural science facts. However, it is generally difficult to fully utilize the historical records, since most of the documents are not written in a modern language and part of the contents are damaged over time. As a result, restoring the damaged or unrecognizable parts as well as translating the records into modern languages are crucial tasks. In response, we present a multi-task learning approach to restore and translate historical documents based on a self-attention mechanism, specifically utilizing two Korean historical records, ones of the most voluminous historical records in the world. Experimental results show that our approach significantly improves the accuracy of the translation task than baselines without multi-task learning. In addition, we present an in-depth exploratory analysis on our translated results via topic modeling, uncovering several significant historical events.

Simultaneous translation, which starts translating each sentence after receiving only a few words in source sentence, has a vital role in many scenarios. Although the previous prefix-to-prefix framework is considered suitable for simultaneous translation and achieves good performance, it still has two inevitable drawbacks: the high computational resource costs caused by the need to train a separate model for each latency $k$ and the insufficient ability to encode information because each target token can only attend to a specific source prefix. We propose a novel framework that adopts a simple but effective decoding strategy which is designed for full-sentence models. Within this framework, training a single full-sentence model can achieve arbitrary given latency and save computational resources. Besides, with the competence of the full-sentence model to encode the whole sentence, our decoding strategy can enhance the information maintained in the decoded states in real time. Experimental results show that our method achieves better translation quality than baselines on 4 directions: Zh$\rightarrow$En, En$\rightarrow$Ro and En$\leftrightarrow$De.