Machine Translation Quality Estimation (QE) is the task of evaluating translation output in the absence of human-written references. Due to the scarcity of human-labeled QE data, previous works attempted to utilize the abundant unlabeled parallel corpora to produce additional training data with pseudo labels. In this paper, we demonstrate a significant gap between parallel data and real QE data: for QE data, it is strictly guaranteed that the source side is original texts and the target side is translated (namely translationese). However, for parallel data, it is indiscriminate and the translationese may occur on either source or target side. We compare the impact of parallel data with different translation directions in QE data augmentation, and find that using the source-original part of parallel corpus consistently outperforms its target-original counterpart. Moreover, since the WMT corpus lacks direction information for each parallel sentence, we train a classifier to distinguish source- and target-original bitext, and carry out an analysis of their difference in both style and domain. Together, these findings suggest using source-original parallel data for QE data augmentation, which brings a relative improvement of up to 4.0% and 6.4% compared to undifferentiated data on sentence- and word-level QE tasks respectively.

The state-of-the-art language model-based automatic metrics, e.g. BARTScore, benefiting from large-scale contextualized pre-training, have been successfully used in a wide range of natural language generation (NLG) tasks, including machine translation, text summarization, and data-to-text. Recent studies show that considering both major errors (e.g. mistranslated tokens) and minor errors (e.g. imperfections in fluency) can produce high-quality human judgments. This inspires us to approach the final goal of the evaluation metrics (human-like evaluations) by automatic error analysis. To this end, we augment BARTScore by incorporating the human-like error analysis strategies, namely BARTScore++, where the final score consists of both the evaluations of major errors and minor errors. Experimental results show that BARTScore++ can consistently improve the performance of vanilla BARTScore and outperform existing top-scoring metrics in 20 out of 25 test settings. We hope our technique can also be extended to other pre-trained model-based metrics. We will release our code and scripts to facilitate the community.

Deep Neural Networks (DNNs) suffer from domain shift when the test dataset follows a distribution different from the training dataset. Domain generalization aims to tackle this issue by learning a model that can generalize to unseen domains. In this paper, we propose a new approach that aims to explicitly remove domain-specific features for domain generalization. Following this approach, we propose a novel framework called Learning and Removing Domain-specific features for Generalization (LRDG) that learns a domain-invariant model by tactically removing domain-specific features from the input images. Specifically, we design a classifier to effectively learn the domain-specific features for each source domain, respectively. We then develop an encoder-decoder network to map each input image into a new image space where the learned domain-specific features are removed. With the images output by the encoder-decoder network, another classifier is designed to learn the domain-invariant features to conduct image classification. Extensive experiments demonstrate that our framework achieves superior performance compared with state-of-the-art methods.

This technical report briefly describes our JDExplore d-team's Vega v2 submission on the SuperGLUE leaderboard. SuperGLUE is more challenging than the widely used general language understanding evaluation (GLUE) benchmark, containing eight difficult language understanding tasks, including question answering, natural language inference, word sense disambiguation, coreference resolution, and reasoning. [Method] Instead of arbitrarily increasing the size of a pretrained language model (PLM), our aim is to 1) fully extract knowledge from the input pretraining data given a certain parameter budget, e.g., 6B, and 2) effectively transfer this knowledge to downstream tasks. To achieve goal 1), we propose self-evolution learning for PLMs to wisely predict the informative tokens that should be masked, and supervise the masked language modeling (MLM) process with rectified smooth labels. For goal 2), we leverage the prompt transfer technique to improve the low-resource tasks by transferring the knowledge from the foundation model and related downstream tasks to the target task. [Results] According to our submission record (Oct. 2022), with our optimized pretraining and fine-tuning strategies, our 6B Vega method achieved new state-of-the-art performance on 4/8 tasks, sitting atop the SuperGLUE leaderboard on Oct. 8, 2022, with an average score of 91.3.

Simultaneous machine translation (SiMT) is usually done via sequence-level knowledge distillation (Seq-KD) from a full-sentence neural machine translation (NMT) model. However, there is still a significant performance gap between NMT and SiMT. In this work, we propose to leverage monolingual data to improve SiMT, which trains a SiMT student on the combination of bilingual data and external monolingual data distilled by Seq-KD. Preliminary experiments on En-Zh and En-Ja news domain corpora demonstrate that monolingual data can significantly improve translation quality (e.g., +3.15 BLEU on En-Zh). Inspired by the behavior of human simultaneous interpreters, we propose a novel monolingual sampling strategy for SiMT, considering both chunk length and monotonicity. Experimental results show that our sampling strategy consistently outperforms the random sampling strategy (and other conventional typical NMT monolingual sampling strategies) by avoiding the key problem of SiMT -- hallucination, and has better scalability. We achieve +0.72 BLEU improvements on average against random sampling on En-Zh and En-Ja. Data and codes can be found at https://github.com/hexuandeng/Mono4SiMT.

Dynamic networks have been extensively explored as they can considerably improve the model's representation power with acceptable computational cost. The common practice in implementing dynamic networks is to convert given static layers into fully dynamic ones where all parameters are dynamic and vary with the input. Recent studies empirically show the trend that the more dynamic layers contribute to ever-increasing performance. However, such a fully dynamic setting 1) may cause redundant parameters and high deployment costs, limiting the applicability of dynamic networks to a broader range of tasks and models, and more importantly, 2) contradicts the previous discovery in the human brain that \textit{when human brains process an attention-demanding task, only partial neurons in the task-specific areas are activated by the input, while the rest neurons leave in a baseline state.} Critically, there is no effort to understand and resolve the above contradictory finding, leaving the primal question -- to make the computational parameters fully dynamic or not? -- unanswered. The main contributions of our work are challenging the basic commonsense in dynamic networks, and, proposing and validating the \textsc{cherry hypothesis} -- \textit{A fully dynamic network contains a subset of dynamic parameters that when transforming other dynamic parameters into static ones, can maintain or even exceed the performance of the original network.} Technically, we propose a brain-inspired partially dynamic network, namely PAD-Net, to transform the redundant dynamic parameters into static ones. Also, we further design Iterative Mode Partition to partition the dynamic- and static-subnet, which alleviates the redundancy in traditional fully dynamic networks. Our hypothesis and method are comprehensively supported by large-scale experiments with typical advanced dynamic methods.

Adapter Tuning, which freezes the pretrained language models (PLMs) and only fine-tunes a few extra modules, becomes an appealing efficient alternative to the full model fine-tuning. Although computationally efficient, the recent Adapters often increase parameters (e.g. bottleneck dimension) for matching the performance of full model fine-tuning, which we argue goes against their original intention. In this work, we re-examine the parameter-efficiency of Adapters through the lens of network pruning (we name such plug-in concept as \texttt{SparseAdapter}) and find that SparseAdapter can achieve comparable or better performance than standard Adapters when the sparse ratio reaches up to 80\%. Based on our findings, we introduce an easy but effective setting ``\textit{Large-Sparse}'' to improve the model capacity of Adapters under the same parameter budget. Experiments on five competitive Adapters upon three advanced PLMs show that with proper sparse method (e.g. SNIP) and ratio (e.g. 40\%) SparseAdapter can consistently outperform their corresponding counterpart. Encouragingly, with the \textit{Large-Sparse} setting, we can obtain further appealing gains, even outperforming the full fine-tuning by a large margin. Our code will be released at: https://github.com/Shwai-He/SparseAdapter.

我们描述了JD Explore Academy对WMT 2022共享的一般翻译任务的提交。我们参加了所有高资源曲目和一条中型曲目，包括中文英语，德语英语，捷克语英语，俄语 - 英语和日语英语。我们通过扩大两个主要因素，即语言对和模型大小，即\ textbf {vega-mt}系统来推动以前的工作的极限 - 进行翻译的双向培训。至于语言对，我们将“双向”扩展到“多向”设置，涵盖所有参与语言，以利用跨语言的常识，并将其转移到下游双语任务中。至于型号尺寸，我们将变压器限制到拥有近47亿参数的极大模型，以完全增强我们VEGA-MT的模型容量。此外，我们采用数据增强策略，例如单语数据的循环翻译以及双语和单语数据的双向自我训练，以全面利用双语和单语言数据。为了使我们的Vega-MT适应通用域测试集，设计了概括调整。根据受约束系统的官方自动分数，根据图1所示的sacrebleu，我们在{zh-en（33.5），en-zh（49.7）（49.7），de-en（33.7）上获得了第一名-de（37.8），CS-EN（54.9），En-CS（41.4）和En-Ru（32.7）}，在{ru-en（45.1）和Ja-en（25.6）}和第三名上的第二名和第三名在{en-ja（41.5）}上； W.R.T彗星，我们在{zh-en（45.1），en-zh（61.7），de-en（58.0），en-de（63.2），cs-en（74.7），ru-en（ru-en（ru-en）上，我们获得了第一名64.9），en-ru（69.6）和en-ja（65.1）}，分别在{en-cs（95.3）和ja-en（40.6）}上的第二名。将发布模型，以通过GitHub和Omniforce平台来促进MT社区。

对抗训练（AT）方法有效地防止对抗性攻击，但它们在不同阶级之间引入了严重的准确性和鲁棒性差异，称为强大的公平性问题。以前建议的公平健壮的学习（FRL）适应重新重量不同的类别以提高公平性。但是，表现良好的班级的表现降低了，导致表现强劲。在本文中，我们在对抗训练中观察到了两种不公平现象：在产生每个类别的对抗性示例（源级公平）和产生对抗性示例时（目标级公平）时产生对抗性示例的不​​同困难。从观察结果中，我们提出平衡对抗训练（BAT）来解决强大的公平问题。关于源阶级的公平性，我们调整了每个班级的攻击强度和困难，以在决策边界附近生成样本，以便更容易，更公平的模型学习；考虑到目标级公平，通过引入统一的分布约束，我们鼓励每个班级的对抗性示例生成过程都有公平的趋势。在多个数据集（CIFAR-10，CIFAR-100和IMAGENETTE）上进行的广泛实验表明，我们的方法可以显着超过其他基线，以减轻健壮的公平性问题（最坏的类精度为+5-10 \％）

大量证据表明，深神经网络（DNN）容易受到后门攻击的影响，这激发了后门检测方法的发展。现有的后门检测方法通常是针对具有单个特定类型（例如基于补丁或基于扰动）的后门攻击而定制的。但是，在实践中，对手可能会产生多种类型的后门攻击，这挑战了当前的检测策略。基于以下事实：对抗性扰动与触发模式高度相关，本文提出了自适应扰动生成（APG）框架，以通过自适应注射对抗性扰动来检测多种类型的后门攻击。由于不同的触发模式在相同的对抗扰动下显示出高度多样的行为，因此我们首先设计了全球到本地策略，以通过调整攻击的区域和预算来适应多种类型的后门触发器。为了进一步提高扰动注入的效率，我们引入了梯度引导的掩模生成策略，以寻找最佳区域以进行对抗攻击。在多个数据集（CIFAR-10，GTSRB，Tiny-Imagenet）上进行的广泛实验表明，我们的方法以大幅度优于最先进的基线（+12％）。