We explore the application of very deep Transformer models for Neural Machine Translation (NMT). Using a simple yet effective initialization technique that stabilizes training, we show that it is feasible to build standard Transformer-based models with up to 60 encoder layers and 12 decoder layers. These deep models outperform their baseline 6-layer counterparts by as much as 2.5 BLEU, and achieve new state-of-the-art benchmark results on WMT14 English-French (43.8 BLEU) and WMT14 English-German (30.1 BLEU).The code and trained models will be publicly available at: https://github.com/namisan/exdeep-nmt.
Learning to rank is an important task that has been successfully deployed in many real-world information retrieval systems. Most existing methods compute relevance judgments of documents independently, without holistically considering the entire set of competing documents. In this paper, we explore modeling documents interactions with self-attention based neural networks. Although self-attention networks have achieved state-of-the-art results in many NLP tasks, we find empirically that self-attention provides little benefit over baseline neural learning to rank architecture. To improve the learning of self-attention weights, We propose simple yet effective regularization terms designed to model interactions between documents. Evaluations on publicly available Learning to Rank (LETOR) datasets show that training self-attention network with our proposed regularization terms can significantly outperform existing learning to rank methods.
We present ESPnet-ST, which is designed for the quick development of speech-to-speech translation systems in a single framework. ESPnet-ST is a new project inside end-to-end speech processing toolkit, ESPnet, which integrates or newly implements automatic speech recognition, machine translation, and text-to-speech functions for speech translation. We provide all-in-one recipes including data pre-processing, feature extraction, training, and decoding pipelines for a wide range of benchmark datasets. Our reproducible results can match or even outperform the current state-of-the-art performances; these pre-trained models are downloadable. The toolkit is publicly available at https://github.com/espnet/espnet.
Despite the reported success of unsupervised machine translation (MT), the field has yet to examine the conditions under which these methods succeed, and where they fail. We conduct an extensive empirical evaluation of unsupervised MT using dissimilar language pairs, dissimilar domains, diverse datasets, and authentic low-resource languages. We find that performance rapidly deteriorates when source and target corpora are from different domains, and that random word embedding initialization can dramatically affect downstream translation performance. We additionally find that unsupervised MT performance declines when source and target languages use different scripts, and observe very poor performance on authentic low-resource language pairs. We advocate for extensive empirical evaluation of unsupervised MT systems to highlight failure points and encourage continued research on the most promising paradigms.
We explore best practices for training small, memory efficient machine translation models with sequence-level knowledge distillation in the domain adaptation setting. While both domain adaptation and knowledge distillation are widely-used, their interaction remains little understood. Our large-scale empirical results in machine translation (on three language pairs with three domains each) suggest distilling twice for best performance: once using general-domain data and again using in-domain data with an adapted teacher.
Adapting machine translation systems in the real world is a difficult problem. In contrast to offline training, users cannot provide the type of fine-grained feedback typically used for improving the system. Moreover, users have different translation needs, and even a single user's needs may change over time. In this work we take a different approach, treating the problem of adapting as one of selection. Instead of adapting a single system, we train many translation systems using different architectures and data partitions. Using bandit learning techniques on simulated user feedback, we learn a policy to choose which system to use for a particular translation task. We show that our approach can (1) quickly adapt to address domain changes in translation tasks, (2) outperform the single best system in mixed-domain translation tasks, and (3) make effective instance-specific decisions when using contextual bandit strategies.
Universal feature extractors, such as BERT for natural language processing and VGG for computer vision, have become effective methods for improving deep learning models without requiring more labeled data. A common paradigm is to pre-train a feature extractor on large amounts of data then fine-tune it as part of a deep learning model on some downstream task (i.e. transfer learning). While effective, feature extractors like BERT may be prohibitively large for some deployment scenarios. We explore weight pruning for BERT and ask: how does compression during pre-training affect transfer learning? We find that pruning affects transfer learning in three broad regimes. Low levels of pruning (30-40\%) do not affect pre-training loss or transfer to downstream tasks at all. Medium levels of pruning increase the pre-training loss and prevent useful pre-training information from being transferred to downstream tasks. High levels of pruning additionally prevent models from fitting downstream datasets, leading to further degradation. Finally, we observe that fine-tuning BERT on a specific task does not improve its prunability. We conclude that BERT can be pruned once during pre-training rather than separately for each task without affecting performance.
Sequence-level knowledge distillation (SLKD) is a model compression technique that leverages large, accurate teacher models to train smaller, under-parameterized student models. Why does pre-processing MT data with SLKD help us train smaller models? We test the common hypothesis that SLKD addresses a capacity deficiency in students by "simplifying" noisy data points and find it unlikely in our case. Models trained on concatenations of original and "simplified" datasets generalize just as well as baseline SLKD. We then propose an alternative hypothesis under the lens of data augmentation and regularization. We try various augmentation strategies and observe that dropout regularization can become unnecessary. Our methods achieve BLEU gains of 0.7-1.2 on TED Talks.
In this paper, we propose a simple yet effective framework for multilingual end-to-end speech translation (ST), in which speech utterances in source languages are directly translated to the desired target languages with a universal sequence-to-sequence architecture. While multilingual models have shown to be useful for automatic speech recognition (ASR) and machine translation (MT), this is the first time they are applied to the end-to-end ST problem. We show the effectiveness of multilingual end-to-end ST in two scenarios: one-to-many and many-to-many translations with publicly available data. We experimentally confirm that multilingual end-to-end ST models significantly outperform bilingual ones in both scenarios. The generalization of multilingual training is also evaluated in a transfer learning scenario to a very low-resource language pair. All of our codes and the database are publicly available to encourage further research in this emergent multilingual ST topic.