There have been significant efforts to interpret the encoder of Transformer-based encoder-decoder architectures for neural machine translation (NMT); meanwhile, the decoder remains largely unexamined despite its critical role. During translation, the decoder must predict output tokens by considering both the source-language text from the encoder and the target-language prefix produced in previous steps. In this work, we study how Transformer-based decoders leverage information from the source and target languages -- developing a universal probe task to assess how information is propagated through each module of each decoder layer. We perform extensive experiments on three major translation datasets (WMT En-De, En-Fr, and En-Zh). Our analysis provides insight on when and where decoders leverage different sources. Based on these insights, we demonstrate that the residual feed-forward module in each Transformer decoder layer can be dropped with minimal loss of performance -- a significant reduction in computation and number of parameters, and consequently a significant boost to both training and inference speed.

We propose an efficient batching strategy for variable-length decoding on GPU architectures. During decoding, when candidates terminate or are pruned according to heuristics, our streaming approach periodically "refills" the batch before proceeding with a selected subset of candidates. We apply our method to variable-width beam search on a state-of-the-art machine translation model. Our method decreases runtime by up to 71% compared to a fixed-width beam search baseline and 17% compared to a variable-width baseline, while matching baselines' BLEU. Finally, experiments show that our method can speed up decoding in other domains, such as semantic and syntactic parsing.

Recent works in Grammatical Error Correction (GEC) have leveraged the progress in Neural Machine Translation (NMT), to learn rewrites from parallel corpora of grammatically incorrect and corrected sentences, achieving state-of-the-art results. At the same time, Generative Adversarial Networks (GANs) have been successful in generating realistic texts across many different tasks by learning to directly minimize the difference between human-generated and synthetic text. In this work, we present an adversarial learning approach to GEC, using the generator-discriminator framework. The generator is a Transformer model, trained to produce grammatically correct sentences given grammatically incorrect ones. The discriminator is a sentence-pair classification model, trained to judge a given pair of grammatically incorrect-correct sentences on the quality of grammatical correction. We pre-train both the discriminator and the generator on parallel texts and then fine-tune them further using a policy gradient method that assigns high rewards to sentences which could be true corrections of the grammatically incorrect text. Experimental results on FCE, CoNLL-14, and BEA-19 datasets show that Adversarial-GEC can achieve competitive GEC quality compared to NMT-based baselines.

Neural autoregressive sequence models are used to generate sequences in a variety of natural language processing (NLP) tasks, where they are evaluated according to sequence-level task losses. These models are typically trained with maximum likelihood estimation, which ignores the task loss, yet empirically performs well as a surrogate objective. Typical approaches to directly optimizing the task loss such as policy gradient and minimum risk training are based around sampling in the sequence space to obtain candidate update directions that are scored based on the loss of a single sequence. In this paper, we develop an alternative method based on random search in the parameter space that leverages access to the maximum likelihood gradient. We propose maximum likelihood guided parameter search (MGS), which samples from a distribution over update directions that is a mixture of random search around the current parameters and around the maximum likelihood gradient, with each direction weighted by its improvement in the task loss. MGS shifts sampling to the parameter space, and scores candidates using losses that are pooled from multiple sequences. Our experiments show that MGS is capable of optimizing sequence-level losses, with substantial reductions in repetition and non-termination in sequence completion, and similar improvements to those of minimum risk training in machine translation.

The lack or absence of parallel and comparable corpora makes bilingual lexicon extraction a difficult task for low-resource languages. The pivot language and cognate recognition approaches have been proven useful for inducing bilingual lexicons for such languages. We propose constraint-based bilingual lexicon induction for closely-related languages by extending constraints from the recent pivot-based induction technique and further enabling multiple symmetry assumption cycles to reach many more cognates in the transgraph. We further identify cognate synonyms to obtain many-to-many translation pairs. This paper utilizes four datasets: one Austronesian low-resource language and three Indo-European high-resource languages. We use three constraint-based methods from our previous work, the Inverse Consultation method and translation pairs generated from the Cartesian product of input dictionaries as baselines. We evaluate our result using the metrics of precision, recall and F-score. Our customizable approach allows the user to conduct cross-validation to predict the optimal hyperparameters (cognate threshold and cognate synonym threshold) with various combinations of heuristics and the number of symmetry assumption cycles to gain the highest F-score. Our proposed methods have statistically significant improvement of precision and F-score compared to our previous constraint-based methods. The results show that our method demonstrates the potential to complement other bilingual dictionary creation methods like word alignment models using parallel corpora for high-resource languages while well handling low-resource languages.

We participate in the WMT 2020 shared news translation task on Chinese to English. Our system is based on the Transformer (Vaswani et al., 2017a) with effective variants and the DTMT (Meng and Zhang, 2019) architecture. In our experiments, we employ data selection, several synthetic data generation approaches (i.e., back-translation, knowledge distillation, and iterative in-domain knowledge transfer), advanced finetuning approaches and self-bleu based model ensemble. Our constrained Chinese to English system achieves 36.9 case-sensitive BLEU score, which is the highest among all submissions.

Models that perform well on a training domain often fail to generalize to out-of-domain (OOD) examples. Data augmentation is a common method used to prevent overfitting and improve OOD generalization. However, in natural language, it is difficult to generate new examples that stay on the underlying data manifold. We introduce SSMBA, a data augmentation method for generating synthetic training examples by using a pair of corruption and reconstruction functions to move randomly on a data manifold. We investigate the use of SSMBA in the natural language domain, leveraging the manifold assumption to reconstruct corrupted text with masked language models. In experiments on robustness benchmarks across 3 tasks and 9 datasets, SSMBA consistently outperforms existing data augmentation methods and baseline models on both in-domain and OOD data, achieving gains of 0.8% accuracy on OOD Amazon reviews, 1.8% accuracy on OOD MNLI, and 1.4 BLEU on in-domain IWSLT14 German-English.

Task-agnostic forms of data augmentation have proven widely effective in computer vision, even on pretrained models. In NLP similar results are reported most commonly for low data regimes, non-pretrained models, or situationally for pretrained models. In this paper we ask how effective these techniques really are when applied to pretrained transformers. Using two popular varieties of task-agnostic data augmentation (not tailored to any particular task), Easy Data Augmentation (Wei and Zou, 2019) and Back-Translation (Sennrichet al., 2015), we conduct a systematic examination of their effects across 5 classification tasks, 6 datasets, and 3 variants of modern pretrained transformers, including BERT, XLNet, and RoBERTa. We observe a negative result, finding that techniques which previously reported strong improvements for non-pretrained models fail to consistently improve performance for pretrained transformers, even when training data is limited. We hope this empirical analysis helps inform practitioners where data augmentation techniques may confer improvements.

Migrating a codebase from an archaic programming language such as COBOL to a modern alternative like Java or C is a difficult, resource-intensive task that requires expertise in both the source and target languages. COBOL, for example, is still widely used today in mainframe systems around the world, so companies, governments, and others often must choose whether to manually translate their code bases or commit to maintaining code written in a language that dates back to the 1950s. We've developed and open sourced TransCoder, an entirely self-supervised neural transcompiler system that can make code migration far easier and more efficient. Our method is the first AI system able to translate code from one programming language to another without requiring parallel data for training. We've demonstrated that TransCoder can successfully translate functions between C, Java, and Python 3. TransCoder outperforms open source and commercial rule-based translation programs.

Retrosynthesis is a problem to infer reactant compounds to synthesize a given product compound through chemical reactions. Recent studies on retrosynthesis focus on proposing more sophisticated prediction models, but the dataset to feed the models also plays an essential role in achieving the best generalizing models. Generally, a dataset that is best suited for a specific task tends to be small. In such a case, it is the standard solution to transfer knowledge from a large or clean dataset in the same domain. In this paper, we conduct a systematic and intensive examination of data transfer approaches on end-to-end generative models, in application to retrosynthesis. Experimental results show that typical data transfer methods can improve test prediction scores of an off-the-shelf Transformer baseline model. Especially, the pre-training plus fine-tuning approach boosts the accuracy scores of the baseline, achieving the new state-of-the-art. In addition, we conduct a manual inspection for the erroneous prediction results. The inspection shows that the pre-training plus fine-tuning models can generate chemically appropriate or sensible proposals in almost all cases.