The deployment of widely used Transformer architecture is challenging because of heavy computation load and memory overhead during inference, especially when the target device is limited in computational resources such as mobile or edge devices. Quantization is an effective technique to address such challenges. Our analysis shows that for a given number of quantization bits, each block of Transformer contributes to translation quality and inference computations in different manners. Moreover, even inside an embedding block, each word presents vastly different contributions. Correspondingly, we propose a mixed precision quantization strategy to represent Transformer weights by an extremely low number of bits (e.g., under 3 bits). For example, for each word in an embedding block, we assign different quantization bits based on statistical property. Our quantized Transformer model achieves 11.8$\times$ smaller model size than the baseline model, with less than -0.5 BLEU. We achieve 8.3$\times$ reduction in run-time memory footprints and 3.5$\times$ speed up (Galaxy N10+) such that our proposed compression strategy enables efficient implementation for on-device NMT.

Recent studies have proven that the training of neural machine translation (NMT) can be facilitated by mimicking the learning process of humans. Nevertheless, achievements of such kind of curriculum learning rely on the quality of artificial schedule drawn up with the handcrafted features, e.g. sentence length or word rarity. We ameliorate this procedure with a more flexible manner by proposing self-paced learning, where NMT model is allowed to 1) automatically quantify the learning confidence over training examples; and 2) flexibly govern its learning via regulating the loss in each iteration step. Experimental results over multiple translation tasks demonstrate that the proposed model yields better performance than strong baselines and those models trained with human-designed curricula on both translation quality and convergence speed.

Recent work on the lottery ticket hypothesis has produced highly sparse Transformers for NMT while maintaining BLEU. However, it is unclear how such pruning techniques affect a model's learned representations. By probing Transformers with more and more low-magnitude weights pruned away, we find that complex semantic information is first to be degraded. Analysis of internal activations reveals that higher layers diverge most over the course of pruning, gradually becoming less complex than their dense counterparts. Meanwhile, early layers of sparse models begin to perform more encoding. Attention mechanisms remain remarkably consistent as sparsity increases.

With the increase in the number of open repositories and discussion forums, the use of natural language for semantic code search has become increasingly common. The accuracy of the results returned by such systems, however, can be low due to 1) limited shared vocabulary between code and user query and 2) inadequate semantic understanding of user query and its relation to code syntax. Siamese networks are well suited to learning such joint relations between data, but have not been explored in the context of code search. In this work, we evaluate Siamese networks for this task by exploring multiple extraction network architectures. These networks independently process code and text descriptions before passing them to a Siamese network to learn embeddings in a common space. We experiment on two different datasets and discover that Siamese networks can act as strong regularizers on networks that extract rich information from code and text, which in turn helps achieve impressive performance on code search beating previous baselines on $2$ programming languages. We also analyze the embedding space of these networks and provide directions to fully leverage the power of Siamese networks for semantic code search.

Despite advances in neural machine translation (NMT) quality, rare words continue to be problematic. For humans, the solution to the rare-word problem has long been dictionaries, but dictionaries cannot be straightforwardly incorporated into NMT. In this paper, we describe a new method for "attaching" dictionary definitions to rare words so that the network can learn the best way to use them. We demonstrate improvements of up to 3.1 BLEU using bilingual dictionaries and up to 0.7 BLEU using monolingual source-language dictionaries.

Sign language translation (SLT) aims to interpret sign video sequences into textbased natural language sentences. Sign videos consist of continuous sequences of sign gestures with no clear boundaries in between. Existing SLT models usually represent sign visual features in a frame-wise manner so as to avoid needing to explicitly segmenting the videos into isolated signs. However, these methods neglect the temporal information of signs and lead to substantial ambiguity in translation. In this paper, we explore the temporal semantic structures of sign videos to learn more discriminative features. To this end, we first present a novel sign video segment representation which takes into account multiple temporal granularities, thus alleviating the need for accurate video segmentation. Taking advantage of the proposed segment representation, we develop a novel hierarchical sign video feature learning method via a temporal semantic pyramid network, called TSPNet. Specifically, TSPNet introduces an inter-scale attention to evaluate and enhance local semantic consistency of sign segments and an intra-scale attention to resolve semantic ambiguity by using non-local video context. Experiments show that our TSPNet outperforms the state-of-the-art with significant improvements on the BLEU score (from 9.58 to 13.41) and ROUGE score (from 31.80 to 34.96) on the largest commonly-used SLT dataset.

As machine translation (MT) systems progress at a rapid pace, questions of their adequacy linger. In this study we focus on negation, a universal, core property of human language that significantly affects the semantics of an utterance. We investigate whether translating negation is an issue for modern MT systems using 17 translation directions as test bed. Through thorough analysis, we find that indeed the presence of negation can significantly impact downstream quality, in some cases resulting in quality reductions of more than 60%. We also provide a linguistically motivated analysis that directly explains the majority of our findings. We release our annotations and code to replicate our analysis here: https://github.com/mosharafhossain/negation-mt.

Lexically cohesive translations preserve consistency in word choices in document-level translation. We employ a copy mechanism into a context-aware neural machine translation model to allow copying words from previous translation outputs. Different from previous context-aware neural machine translation models that handle all the discourse phenomena implicitly, our model explicitly addresses the lexical cohesion problem by boosting the probabilities to output words consistently. We conduct experiments on Japanese to English translation using an evaluation dataset for discourse translation. The results showed that the proposed model significantly improved lexical cohesion compared to previous context-aware models.

Transformers are being used extensively across several sequence modeling tasks. Significant research effort has been devoted to experimentally probe the inner workings of Transformers. However, our conceptual and theoretical understanding of their power and inherent limitations is still nascent. In particular, the roles of various components in Transformers such as positional encodings, attention heads, residual connections, and feedforward networks, are not clear. In this paper, we take a step towards answering these questions. We analyze the computational power as captured by Turing-completeness. We first provide an alternate and simpler proof to show that vanilla Transformers are Turing-complete and then we prove that Transformers with only positional masking and without any positional encoding are also Turing-complete. We further analyze the necessity of each component for the Turing-completeness of the network; interestingly, we find that a particular type of residual connection is necessary. We demonstrate the practical implications of our results via experiments on machine translation and synthetic tasks.

Open-domain response generation is the task of generating sensible and informative re-sponses to the source sentence. However, neural models tend to generate safe and mean-ingless responses. While cue-word introducing approaches encourage responses with concrete semantics and have shown tremendous potential, they still fail to explore di-verse responses during decoding. In this paper, we propose a novel but natural approach that can produce multiple cue-words during decoding, and then uses the produced cue-words to drive decoding and shrinks the decoding vocabulary. Thus the neural genera-tion model can explore the full space of responses and discover informative ones with efficiency. Experimental results show that our approach significantly outperforms several strong baseline models with much lower decoding complexity. Especially, our approach can converge to concrete semantics more efficiently during decoding.