1 Introduction
Neural machine translation (NMT) has shown promising results and drawn more attention recently Kalchbrenner and Blunsom (2013); Cho et al. (2014b); Bahdanau et al. (2015); Gehring et al. (2017a, b); Vaswani et al. (2017). A widely used architecture is the attentionbased encoderdecoder framework Cho et al. (2014b); Bahdanau et al. (2015)
which assumes there is a common semantic space between the source and target language pairs. The encoder encodes the source sentence to a representation in the common space with the recurrent neural network (RNN)
Hochreiter and Schmidhuber (1997)and the decoder decodes this representation to generate the target sentence word by word. To generate a target word, a probability distribution over the target vocabulary is drawn based on the attention over the entire source sequence and the target information rolled by another RNN. At the training time, the decoder is forced to generate the ground truth sentence, while at inference, it needs to employ the beam search algorithm to search through a constrained space due to the huge search space.
Even with beam search, NMT still suffers from slow translation speed, especially when it works not on GPUs, but on CPUs, which are more common practice. The first reason for the inefficiency is that the generation of each target word requires extensive computation to go through all the source words to calculate the attention. Worse still, due to the recurrence of RNNs, target words can only be generated sequentially rather than in parallel. The second reason is that large vocabulary on target side is employed to avoid unknown words (UNKs), which leads to a large number of normalization factors for the operation when drawing the probability distribution. To accelerate the translation, the widely used method is to trade off between the translation quality and the decoding speed by reducing the size of vocabulary Mi et al. (2016a) or/and the number of parameters, which can not realize the full potential of NMT.
In this paper, we borrow ideas from phrasebased and syntaxbased machine translation where cube pruning has been successfully applied to speed up the decoding Chiang (2007); Huang and Chiang (2007). Informally, cube pruning “coarsens” the search space by clustering similar states according to some equivalence relations. To apply this idea to NMT, however, is much more involved. Specifically, in the process of beam search, we cluster similar target hidden states to construct equivalence classes, the three dimensions of which are target words in the target vocabulary, part translations retained in the beam search and different combinations of similar target hidden states, respectively. The clustering operation can directly decrease the number of target hidden states in the following calculations, together with cube pruning, resulting in less RNN expansion operations to generate the next hidden state (related to the first reason) and less operations over the target vocabulary (related to the second reason). The experiment results show that, when receiving the same or even better translation quality, our method can speed up the decoding speed by on GPUs and on CPUs.
2 Background
The proposed strategy can be adapted to optimize the beam search algorithm in the decoder of various NMT models. Without loss of generality, we take the attentionbased NMT Bahdanau et al. (2015) as an example to introduce our method. In this section, we first introduce the attentionbased NMT model and then the cube pruning algorithm.
2.1 The Attentionbased NMT Model
The attentionbased NMT model follows the encoderdecoder framework with an extra attention module. In the following parts, we will introduce each of the three components. Assume the source sequence and the observed translation are and .
Encoder The encoder uses a bidirectional GRU to obtain two sequences of hidden states. The final hidden state of each source word is got by concatenating the corresponding pair of hidden states in those sequences. Note that
is employed to represent the embedding vector of the word
.(1)  
(2)  
(3) 
Attention The attention module is designed to extract source information (called context vector) which is highly related to the generation of the next target word. At the th step, to get the context vector, the relevance between the target word and the th source word is firstly evaluated as
(4) 
Then, the relevance is normalized over the source sequence, and all source hidden states are added weightedly to produce the context vector.
(5) 
Decoder The decoder also employs a GRU to unroll the target information. The details are described in Bahdanau et al. (2015). At the th decoding step, the target hidden state is given by
(6) 
The probability distribution over all the words in the target vocabulary is predicted conditioned on the previous ground truth words, the context vector and the unrolled target information .
(7)  
(8)  
(9) 
where
stands for a linear transformation,
is used to map to so that each target word has one corresponding dimension in .2.2 Cube Pruning
GPU  CPU  

Calculation Units  Time(s)  Percentage  Time(s)  Percentage 
Eq. (6):  551.07  75.73%  1370.92  19.42% 
Eq. (7):  88.25  12.13%  277.76  3.93% 
Eq. (8):  25.33  3.48%  2342.53  33.18% 
Eq. (9):  63.00  8.66%  3069.25  43.47% 
The cube pruning algorithm, proposed by Chiang (2007) based on the best parsing algorithm of Huang and Chiang (2005)
, is actually an accelerated extension based on the naive beam search algorithm. Beam search, a heuristic dynamic programming searching algorithm, explores a graph by expanding the most promising nodes in a limited set and searches approximate optimal results from candidates. For the sequencetosequence learning task, given a pretrained model, the beam search algorithm finds a sequence that approximately maximizes the conditional probability
Graves (2012); BoulangerLewandowski et al. (2013). Both Sutskever et al. (2014) and Bahdanau et al. (2015) employed the beam search algorithm into the NMT decoding to produce translations with relatively larger conditional probability with respect to the optimized model parameters. Remarkably, Huang and Chiang (2007) successfully applied the cube pruning algorithm to the decoding of SMT. They found that the beam search algorithm in SMT can be extended, and they utilized the cube pruning and some variants to optimize the search process in the decoding phase of phrasebased Och and Ney (2004) and syntaxbased Chiang (2005); Galley et al. (2006) systems, which decreased a mass of translation candidates and achieved a significant speed improvement by reducing the size of complicated search space, thereby making it possible to actualize the thought of improving the translation performance through increasing the beam size.In the traditional SMT decoding, the cube pruning algorithm aims to prune a great number of partial translation hypotheses without computing and storing them. For each decoding step, those hypotheses with the same translation rule are grouped together, then the cube pruning algorithm is conducted over the hypotheses. We illustrate the detailed process in Figure 1.
3 NMT Decoder with Cube Pruning
3.1 Definitions
We define the related storage unit tuple of the th candidate word in the th beam as , where is the negative loglikelihood (NLL) accumulation in the th beam, is the decoder hidden state in the th beam, is the index of the th target word in large vocabulary and is the backtracking pointer for the th decoding step. Note that, for each source sentence, we begin with calculating its encoded representation and the first hidden state in decoder, then searching from the initial tuple existing in the first beam^{1}^{1}1The initial target word index equals to , which actually corresponds to the Beginning Of Sentence (BOS) token in target vocabulary..
It is a fact that Equation (9) produces the probability distribution of the predicted target words over the target vocabulary . Cho et al. (2014b) indicated that whenever a target word is generated, the function over computes probabilities for all words in , so the calculation is expensive when the target vocabulary is large. As such, Bahdanau et al. (2015) (and many others) only used the top frequent words as target vocabulary, and replaced others with UNK. However, the final normalization operation still brought high computation complexity for forward calculations.
3.2 Time Cost in Decoding
We conducted an experiment to explore how long each calculation unit in the decoder would take. We decoded the MT03 test dataset by using naive beam search with beam size of and recorded the time consumed in the computation of Equation (6), (7), (8) and (9), respectively. The statistical results in Table 1 show that the recurrent calculation unit consumed the most time on GPUs, while the computation also took lots of time. On CPUs, the most expensive computational time cost was caused by the operation over the entire target vocabulary^{2}^{2}2Note that, identical to Bahdanau et al. (2015), we only used as the vocabulary size.. In order to avoid the timeconsuming normalization operation in testing, we introduced selfnormalization (denoted as SN) into the training.
3.3 Selfnormalization
Selfnormalization Devlin et al. (2014) was designed to make the model scores which are produced by the output layer be approximated by the probability distribution over the target vocabulary without normalization operation. According to Equation (9), for an observed target sentence , the CrossEntropy (CE) loss could be written as
(10)  
where is the model score generated by Equation (8) at the th step, we marked the normalizer as .
The objective function, shown in Equation (11), is optimized to make sure is approximated to , equally, make close to once it converges. We chose the value of empirically. Because the normalizer is converged to in inference, we just need to ignore and predict the target word distribution at the th step only with :
(12) 
3.4 Cube Pruning
Table 1 clearly shows that the equations in the NMT forward calculation take lots of time. Here, according to the idea behind the cube pruning algorithm, we tried to reduce the time of timeconsuming calculations, e.g., Equation (6), and further decrease the search space by introducing the cube pruning algorithm.
3.4.1 Integrating into NMT Decoder
Extended from the naive beam search in the NMT decoder, cube pruning, treated as a pruning algorithm, attempts to reduce the search space and computation complexity by merging some similar items in a beam to accelerate the naive beam search, keeping the best searching result almost unchanged or even better by increasing the beam size. Thus, it is a fast and effective algorithm to generate candidates.
Assume that restores the set of the finished translations. For each step in naive beam search process, times forward calculations are required to acquire probability distributions corresponding to each item in the previous beam Bahdanau et al. (2015). while for each step in cube pruning, in terms of some constraints, we merge all similar items in the previous beam into one equivalence class (called a subcube). The constraint we used here is that items being merged in the previous beam should have the same target words. Then, for the subcube, only one forward calculation is required to obtain the approximate predictions by using the loose hidden state. Elements in the subcube are sorted by previous accumulated NLL along the columns (the first dimension of beam size) and by the approximate predictions along the rows (the second dimension of vocabulary size). After merging, one beam may contain several subcubes (the third dimension), we start to search from item in the upper left corner of each subcube, which is the best one in the subcube, and continue to spread out until enough candidates are found. Once a item is selected, the exact hidden state will be used to calculate its exact NLL.
Through all above steps, the frequency of forward computations decreases. We give an example to dive into the details in Figure 2.
Assume that the beam size is . Given the beam, we generate the beam. Different from the naive beam search, we first group items in the previous beam into two subcubes and in term of the target word . As shown in part of Figure 2, constructs the subcube ; , and are put together to compose another subcube . Items in part are ranked in ascending order along both row and column dimension according to the accumulated NLL. For each subcube, we use the first state vector in each subcube as the approximate one to produce the next probability distribution and the next state. At beginning, each upperleft corner element in each subcube is pushed into a minimum heap, after popping minimum element from the heap, we calculate and restore the exact NLL of the element, then push the right and lower ones alongside the minimum element into heap. At this rate, the searching continues just like the “diffusion” in the subcube until elements are popped, which are ranked in terms of their exact NLLs to construct the beam. Note that once an element is popped, we calculate its exact NLL. From the step (e) in Figure 2, we can see that elements have been popped from and , and then ranked in terms of their exact NLLs to build the beam.
We refer above algorithm as the naive cube pruning algorithm (called NCP)
3.4.2 Accelerated Cube Pruning
In each step of the cube pruning algorithm, after merging the items in the previous beam, some similar candidates are grouped together into one or more subcube(s). We also try to predict the approximate distribution for each subcube only according to the top state vector (the first row in the subcube in Figure 2), and select next candidates after ranking. The predicted probability distribution will be very similar to that of the naive beam search. Besides, Each subcube only requires one forward calculation. Thus, it could save more search space and further reduce the computation complexity for the decoder. Unlike the naive cube pruning algorithm, accelerated cube pruning pops each item, then still use the approximate NLL instead of the exact one. We denote this kind of accelerated cube pruning algorithm as ACP.
4 Experiments
We verified the effectiveness of proposed cube pruning algorithm on the ChinesetoEnglish (ZhEn) translation task.
4.1 Data Preparation
The ChineseEnglish training dataset consists of M sentence pairs^{3}^{3}3These sentence pairs are mainly extracted from LDC2002E18, LDC2003E07, LDC2003E14, Hansards portion of LDC2004T07, LDC2004T08 and LDC2005T06. We used the NIST 2002 (MT02) dataset as the validation set with sentences, and the NIST 2003 (MT03) dataset as the test dataset, which contains sentences.
The lengths of the sentences on both sides were limited up to tokens, then actually M sentence pairs were left with M Chinese words and M English words. We extracted most frequent words as the source and target vocabularies for both sides.
In all the experiments, caseinsensitive gram BLEU Papineni et al. (2002) was employed for the automatic evaluation, we used the script mtevalv11b.pl^{4}^{4}4https://github.com/mosessmt/mosesdecoder/blob/master/scripts/generic/mtevalv11b.pl to calculate the BLEU score.
4.2 System
The system is an improved version of attentionbased NMT system named RNNsearch Bahdanau et al. (2015) where the decoder employs a conditional GRU layer with attention, consisting of two GRUs and an attention module for each step^{5}^{5}5https://github.com/nyudl/dl4mttutorial/blob/master/docs/cgru.pdf. Specifically, Equation (6) is replaced with the following two equations:
(13)  
(14) 
Besides, for the calculation of relevance in Equation (4), is replaced with . The other components of the system keep the same as RNNsearch. Also, we reimplemented the beam search algorithm as the naive decoding method, and naive searching on the GPU and CPU server were conducted as two baselines.
4.3 Training Details
Specially, we employed a little different settings from Bahdanau et al. (2015): Word embedding sizes on both sides were set to , all hidden sizes in the GRUs of both encoder and decoder were also set to
. All parameter matrices, including bias matrices, were initialized with the uniform distribution over
. Parameters were updated by using minibatch Stochastic Gradient Descent (SGD) with batch size of
and the learning rate was adjusted by AdaDelta Zeiler (2012) with decay constant = and denominator constant =. The gradients of all variables whose norm are larger than a predefined threshold were normalized to the threshold to avoid gradient explosion Pascanu et al. (2013). Dropout was applied to the output layer with dropout rate of . We exploited length normalization Cho et al. (2014a) strategy on candidate translations in beam search decoding.The model whose BLEU score was the highest on the validation set was used to do testing. Maximal epoch number was set to
. Training was conducted on a single Tesla K80 GPU, it took about days to train a single NMT model on the ZhEn training data. For selfnormalization, we empirically set as in Equation (11)^{6}^{6}6Following Devlin et al. (2014), we had tried , , and for the value of , we found that produced the best result..4.4 Search Strategies
We conducted experiments to decode the MT03 test dataset on the GPU and CPU server respectively, then compared search quality and efficiency among following six search strategies under different beam sizes.
NBSSN: Naive Beam Search without SN
NBS+SN: Naive Beam Search with SN
NCPSN: Cube Pruning without SN
NCP+SN: Cube Pruning with SN
ACPSN: Accelerated Cube Pruning without SN
ACP+SN: Accelerated Cube Pruning with SN
4.5 Comparison of Average Merging Rate
We first give the definition of the Average Merging Rate (denoted as AMR). Given a test dataset, we counted the total wordlevel candidates (noted as ) and the total subcubes (noted as ) during the whole decoding process, then the AMR can be simply computed as
(15) 
The MT03 test dataset was utilized to compare the trends of the AMR values under all six methods. We used the pretrained model to translate the test dataset on a single GeForce GTX TITAN X GPU server. Beam size varies from to , values are included in the set . For each beam size, six different searching settings were applied to translate the test dataset respectively. The curves of the AMRs during the decoding on the MT03 test dataset under the proposed methods are shown in Figure 3. Note that the AMR values of NBS are always whether there is SN or not.
Comparing the curves in the Figure 3, we could observe that the naive beam search does not conduct any merging operation in the whole searching process, while the average merging rate in the cube pruning almost grows as the beam size increases. Comparing the red curves to the blue ones, we can conclude that, in any case of beam size, the AMR of the accelerated cube pruning surpasses the basic cube pruning by a large margin. Besides, selfnormalization could produces the higher average merging rate comparing to the counterpart without selfnormalization.
4.6 Comparison on the GPU Server
Intuitively, as the value of the AMR increases, the search space will be reduced and computation efficiency improves. We compare the two proposed searching strategies and the naive beam search in two conditions (with selfnormalization and without selfnormalization). Figure 4 demonstrates the results of comparison between the proposed searching methods and the naive beam search baseline in terms of search quality and search efficiency under different beam sizes.
By fixing the beam size and the dataset, we compared the changing trend of BLEU scores for the three distinct searching strategies under two conditions. Without selfnormalization, Figure 3(a) shows the significant improvement of the search speed, however the BLEU score drops about points. We then equipped the search algorithm with selfnormalization. Figure 3(b) shows that the accelerated cube pruning search algorithm only spend about onethird of the time of the naive beam search to achieve the best BLEU score with beam size . Concretely, when the beam size is set to be , ACP+SN is times faster than the baseline on the MT03 test dataset, and both performances are almost the same.
4.7 Comparison on the CPU Server
Similar to the experiments conducted on GPUs, we also translated the whole MT03 test dataset on the CPU server by using all six search strategies under different beam sizes. The trends of the BLEU scores over those strategies are shown in Figure 5.
The proposed search methods gain the similar superiority on CPUs to that on GPUs, and the decoding speed is obviously slower than that on GPUs. From the Figure 4(a), we can also clearly see that, compared with the NBSSN, NCPSN only speeds up the decoder a little, ACPSN produces much more acceleration. However, when we did not introduce selfnormalization, the proposed search methods will also result in a loss of about BLEU score. The selfnormalization made the ACP strategy faster than the baseline by about , in which condition the NBS+SN got the best BLEU score with beam size while the ACP+SN achieved the highest score with beam size . The results could be observed in Figure 4(b). Because our method is on the algorithmic level and platformindependent, it is reasonable that the proposed method can not only perform well on GPUs, but also accelerate the decoding significantly on CPUs. Thus, the accelerated cube pruning with selfnormalization could improve the search quality and efficiency stably.
4.8 Decoding Time
In this section, we only focus on the consuming time of translating the entire MT03 test dataset. Under the two conditions, we calculated the times spent on translating the entire test dataset for different beam sizes, then draw the curves in Figure 6 and 7. From the Figure 5(a) and 5(b), we could observe that accelerated cube pruning algorithm speeds up the decoding by about on GPUs when the beam size is set to . Figure 6(a) and 6(b) show that the accelerated cube pruning algorithm speeds up the decoding by about on CPU server with the beam size .
5 Related Work
Recently, lots of works devoted to improve the efficiency of the NMT decoder. Some researchers employed the way of decreasing the target vocabulary size. Jean et al. (2015) improved the decoding efficiency even with the model using a very large target vocabulary but selecting only a small subset of the whole target vocabulary. Based on the work of Jean et al. (2015), Mi et al. (2016b) introduced sentencelevel and batchlevel vocabularies as a very small subset of the full output vocabulary, then predicted target words only on this small vocabulary, in this way, they only lost BLEU points, but reduced target vocabulary substantially.
Some other researchers tried to raise the efficiency of decoding from other perspectives. Wu et al. (2016) introduced a coverage penalty and length normalization into beam search decoder to prune hypotheses and sped up the search process by when running on CPUs. Hu et al. (2015) used a priority queue to choose the best hypothesis for the next search step, which drastically reduced search space.
Inspired by the works of Mi et al. (2016b) and Huang and Chiang (2007), we consider pruning hypothesis in NMT decoding by using cube pruning algorithm, but unlike traditional SMT decoding where dynamic programming was used to merge equivalent states (e.g., if we use phrasebased decoding with trigram language model, we can merge states with same sourceside coverage vector and same previous two target words). However, this is not appropriate for current NMT decoding, since the embedding of the previous target word is used as one input of the calculation unit of each step in the decoding process, we could group equivalence classes containing the same previous target word together.
6 Conclusions
We extended cube pruning algorithm into the decoder of the attentionbased NMT. For each step in beam search, we grouped similar candidates in previous beam into one or more equivalence class(es), and bad hypotheses were pruned out. We started searching from the upperleft corner in each equivalence class and spread out until enough candidates were generated. Evaluations show that, compared with naive beam search, our method could improve the search quality and efficiency to a large extent, accelerating the NMT decoder by and on GPUs and CPUs, respectively. Also, the translation precision could be the same or even better in both situations. Besides, selfnormalization is verified to be helpful to accelerate cube pruning even further.
Acknowledgements
We thank the three anonymous reviewers for their comments, Kai Zhao and Haitao Mi for uggestions. This work is supported in part by NSF IIS1817231 & IIS1656051, and is also supported in part by National Natural Science Foundation of China (No. 61472428 & No. 61662077).
References
 Bahdanau et al. (2015) Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. ICLR 2015.
 BoulangerLewandowski et al. (2013) Nicolas BoulangerLewandowski, Yoshua Bengio, and Pascal Vincent. 2013. Audio chord recognition with recurrent neural networks. In ISMIR, pages 335–340. Citeseer.
 Chiang (2005) David Chiang. 2005. A hierarchical phrasebased model for statistical machine translation. In Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics, pages 263–270. Association for Computational Linguistics.
 Chiang (2007) David Chiang. 2007. Hierarchical phrasebased translation. computational linguistics, 33(2):201–228.
 Cho et al. (2014a) Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. 2014a. On the properties of neural machine translation: Encoder–decoder approaches. In Proceedings of SSST8, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, pages 103–111, Doha, Qatar. Association for Computational Linguistics.

Cho et al. (2014b)
Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi
Bougares, Holger Schwenk, and Yoshua Bengio. 2014b.
Learning phrase representations using rnn encoder–decoder for
statistical machine translation.
In
Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)
, pages 1724–1734, Doha, Qatar. Association for Computational Linguistics.  Devlin et al. (2014) Jacob Devlin, Rabih Zbib, Zhongqiang Huang, Thomas Lamar, Richard Schwartz, and John Makhoul. 2014. Fast and robust neural network joint models for statistical machine translation. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1370–1380, Baltimore, Maryland. Association for Computational Linguistics.
 Galley et al. (2006) Michel Galley, Jonathan Graehl, Kevin Knight, Daniel Marcu, Steve DeNeefe, Wei Wang, and Ignacio Thayer. 2006. Scalable inference and training of contextrich syntactic translation models. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics, pages 961–968, Sydney, Australia. Association for Computational Linguistics.
 Gehring et al. (2017a) Jonas Gehring, Michael Auli, David Grangier, and Yann Dauphin. 2017a. A convolutional encoder model for neural machine translation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 123–135, Vancouver, Canada. Association for Computational Linguistics.

Gehring et al. (2017b)
Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin.
2017b.
Convolutional sequence to sequence learning.
In
Proceedings of the 34th International Conference on Machine Learning
, volume 70 of Proceedings of Machine Learning Research, pages 1243–1252, International Convention Centre, Sydney, Australia. PMLR.  Graves (2012) Alex Graves. 2012. Sequence transduction with recurrent neural networks. arXiv preprint arXiv:1211.3711.
 Hochreiter and Schmidhuber (1997) Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long shortterm memory. Neural Computation, 9(8):1735–1780.
 Hu et al. (2015) Xiaoguang Hu, Wei Li, Xiang Lan, Hua Wu, and Haifeng Wang. 2015. Improved beam search with constrained softmax for nmt. Proceedings of MT Summit XV, page 297.
 Huang and Chiang (2005) Liang Huang and David Chiang. 2005. Better kbest parsing. In Proceedings of the Ninth International Workshop on Parsing Technology, Parsing ’05, pages 53–64, Stroudsburg, PA, USA. Association for Computational Linguistics.
 Huang and Chiang (2007) Liang Huang and David Chiang. 2007. Forest rescoring: Faster decoding with integrated language models. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 144–151, Prague, Czech Republic. Association for Computational Linguistics.
 Jean et al. (2015) Sébastien Jean, Kyunghyun Cho, Roland Memisevic, and Yoshua Bengio. 2015. On using very large target vocabulary for neural machine translation. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1–10, Beijing, China. Association for Computational Linguistics.

Kalchbrenner and Blunsom (2013)
Nal Kalchbrenner and Phil Blunsom. 2013.
Recurrent convolutional neural networks for discourse compositionality.
In Proceedings of the Workshop on Continuous Vector Space Models and their Compositionality, pages 119–126, Sofia, Bulgaria. Association for Computational Linguistics.  Mi et al. (2016a) Haitao Mi, Baskaran Sankaran, Zhiguo Wang, and Abe Ittycheriah. 2016a. Coverage embedding models for neural machine translation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 955–960, Austin, Texas. Association for Computational Linguistics.
 Mi et al. (2016b) Haitao Mi, Zhiguo Wang, and Abe Ittycheriah. 2016b. Vocabulary manipulation for neural machine translation. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 124–129, Berlin, Germany. Association for Computational Linguistics.
 Och and Ney (2004) Franz Josef Och and Hermann Ney. 2004. The alignment template approach to statistical machine translation. Computational Linguistics, 30(4):417–449.
 Papineni et al. (2002) Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics.
 Pascanu et al. (2013) Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. 2013. On the difficulty of training recurrent neural networks. In Proceedings of the 30th International Conference on Machine Learning, volume 28 of Proceedings of Machine Learning Research, pages 1310–1318, Atlanta, Georgia, USA. PMLR.
 Sutskever et al. (2014) Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. In Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 27, pages 3104–3112. Curran Associates, Inc.
 Vaswani et al. (2017) Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems 30, pages 5998–6008. Curran Associates, Inc.
 Wu et al. (2016) Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. 2016. Google’s neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144.
 Zeiler (2012) Matthew D Zeiler. 2012. Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701.
Comments
There are no comments yet.