Multilingual Neural Machine Translation for Zero-Resource Languages
In recent years, Neural Machine Translation (NMT) has been shown to be more effective than phrase-based statistical methods, thus quickly becoming the state of the art in machine translation (MT). However, NMT systems are limited in translating low-resourced languages, due to the significant amount of parallel data that is required to learn useful mappings between languages. In this work, we show how the so-called multilingual NMT can help to tackle the challenges associated with low-resourced language translation. The underlying principle of multilingual NMT is to force the creation of hidden representations of words in a shared semantic space across multiple languages, thus enabling a positive parameter transfer across languages. Along this direction, we present multilingual translation experiments with three languages (English, Italian, Romanian) covering six translation directions, utilizing both recurrent neural networks and transformer (or self-attentive) neural networks. We then focus on the zero-shot translation problem, that is how to leverage multi-lingual data in order to learn translation directions that are not covered by the available training material. To this aim, we introduce our recently proposed iterative self-training method, which incrementally improves a multilingual NMT on a zero-shot direction by just relying on monolingual data. Our results on TED talks data show that multilingual NMT outperforms conventional bilingual NMT, that the transformer NMT outperforms recurrent NMT, and that zero-shot NMT outperforms conventional pivoting methods and even matches the performance of a fully-trained bilingual system.
Neural machine translation (NMT) has shown its effectiveness by delivering the best performance in the IWSLT  and WMT  evaluation campaigns of the last three years. Unlike rule-based or statistical MT, the end-to-end learning approach of NMT models the mapping from source to target language directly through a posterior probability. The essential component of an NMT system includes an encoder, a decoder and an attention mechanism . Despite the continuous improvement in performance and translation quality , NMT models are highly dependent on the availability of extensive parallel data, which in practice can only be acquired for a very limited number of language pairs. For this reason, building effective NMT systems for low-resource languages becomes a primary challenge . In fact, \namecitezoph2016transfer showed how a standard string-to-tree statistical MT system  can effectively outperform NMT methods for low-resource languages, such as Hausa, Uzbek, and Urdu.
In this work, we approach low-resource machine translation with so-called multilingual NMT [17, 15], which considers the use of NMT to target many-to-many translation directions. Our motivation is that intensive cross-lingual transfer  via parameter sharing across multiple languages should ideally help in the case of similar languages and sparse training data. In particular, we investigate multilingual NMT across Italian, Romanian, and English, and simulate low-resource conditions by limiting the amount of available parallel data.
Among the various approaches for multilingual NMT, the simplest and most effective one is to train a single neural network on parallel data including multiple translation directions and to prepend to each source sentence a flag specifying the desired target language [15, 17]. In this work, we investigate multi-lingual NMT under low-resource conditions and with two popular NMT architectures: recurrent LSTM-based NMT and the self-attentive or transformer NMT model. In particular, we train and evaluate our systems on a collection of TED talks , over six translation directions: EnglishItalian, EnglishRomanian, and ItalianRomanian.
A major advantage of multi-lingual NMT is the possibility to perform a zero-shot translation, that is to query the system on a direction for which no training data was provided. The case we consider is illustrated in Figure 1: we assume to only have Italian-English and English-Romanian training data and that we need to translate between Italian and Romanian in both directions. To solve this task, we propose a self-learning method that permits a multi-lingual NMT system trained on the above mentioned language pairs to progressively learn how to translate between Romanian and Italian directly from its own translations. We show that our zero-shot self-training approach not only improves over a conventional pivoting approach, by bridging Romanian-Italian through English (see Figure 1), but that it also matches the performance of bilingual systems trained on Italian-Romanian data. The contribution of this work is twofold:111This paper integrates and extends work presented in  and .
Comparing RNN and Transformer approaches in a multilingual NMT setting;
A self-learning approach to improve the zero-shot translation task of a multilingual model.
The paper is organized as follows. In Section 2, we present previous works on multilingual NMT, zero-shot NMT, and NMT training with self-generated data. In Section 3, we introduce the two prominent NMT approaches evaluated in this paper, the recurrent and the transformer models. In Section 4, we introduce our multilingual NMT approach and our self-training method for zero-shot learning. In Section 5, we describe our experimental set-up and the NMT model configurations. In Section 6, we present and discuss the results of our experiments. Section 7 ends the paper with our conclusions.
2 Previous Work
2.1 Multilingual NMT
Previous works in multilingual NMT are characterized by the use of separate encoding and/or decoding networks for each translation direction. \namecitedong2015multi proposed a multi-task learning approach for a one-to-many translation scenario, by sharing hidden representations among related tasks – i.e the source languages – to enhance generalization on the target language. In particular, they used a single encoder for all source languages and separate attention mechanisms and decoders for every target language. In a related work, \nameciteluong2015multi used distinct encoder and decoder networks for modeling language pairs in a many-to-many setting. Aimed at reducing ambiguities at translation time, \namecitezoph2016multi employed a many-to-one system that considers two languages on the encoder side and one target language on the decoder side. In particular, the attention model is applied to a combination of the two encoder states. In a many-to-many translation scenario, \namecitefirat2016multi introduced a way to share the attention mechanism across multiple languages. As in \namecitedong2015multi (but only on the decoder side) and in \nameciteluong2015multi, they used separate encoders and decoders for each source and target language.
Despite the reported improvements, the need of using an additional encoder and/or decoder for every language added to the system tells the limitation of these approaches, by making their network complex and expensive to train.
In a very different way, \namecitejohnson2016google and \nameciteha2016toward developed similar multilingual NMT approaches by introducing a target-forcing token in the input. The approach in \nameciteha2016toward applies a language-specific code to words from different languages in a mixed-language vocabulary. In practice, they force the decoder to translate into a specific target language by prepending and appending an artificial token to the source text. However, their word and sub-word level language-specific coding mechanism significantly increases the input length, which shows to have an impact on the computational cost and performance of NMT . In \namecitejohnson2016google, only one artificial token is pre-pended to the entire source sentences in order to specify the target language. Hence, the same token is also used to trigger the decoder generating the translation (cf. Figure 2). Remarkably, pre-pending language tokens to the input string has greatly simplified multi-lingual NMT, by eliminating the need of having separate encoder/decoder networks and attention mechanism for every new language pair.
2.2 Zero-Shot and Self-Learning
Zero-resource NMT has been proposed in  and it extends the work by \namecitefirat2016multi. The authors proposed a many-to-one translation setting and used the idea of generating a pseudo-parallel corpus , using a pivot language, to fine tune their model. However, also in this case the need of separate encoders and decoders for every language pair significantly increases the complexity of the model.
An attractive feature of the target-forcing mechanism comes from the possibility to perform zero-shot translation with the same multilingual setting as in [17, 15]. Both the works reported that a multilingual system trained on a large amount of data improves over a baseline bilingual model and that it is also capable of performing zero-shot translation, assuming that the zero-shot source and target languages have been observed during training paired with some other languages.
However, recent experiments have shown that the mechanism fails to achieve reasonable zero-shot translation performance for low-resource languages . The promising results in  and  hence require further investigation to verify if their method can work in various language settings, particularly across distant languages.
An alternative approach to zero-shot translation in a resource-scarce scenario is to use a pivot language , that is, using an intermediate language for translation. While this solution is usually pursued by deploying two or more bilingual models, in this work we aim to achieve comparable results using a single multilingual model.
Training procedures using synthetic data have been around for a while. For instance, in statistical machine translation (SMT), \nameciteoflazer2007exploringIncremental and \namecitebechara2011statisticalIncremental showed how the output of a translation model can be used iteratively to improve results in a task like post-editing. Mechanisms like back-translating the target side of a single language pair have been used for domain adaptation  and more recently by \namecitesennrich2015improvingMono to improve an NMT baseline model. In , a dual-learning mechanism is proposed where two NMT models working in the opposite directions provide each other feedback signals that permit them to learn from monolingual data. In a related way, our approach also considers training from monolingual data. As a difference, however, our proposed method leverages the capability of the network to jointly learn multiple translation directions and to directly generate the translations used for self-training.
Although our brief survey shows that re-using the output of an MT system for further training and improvement has been successfully applied in different settings, our approach differs from past works in two aspects: i) introducing a new self-training method integrated in a multilingual NMT architecture, and ii) casting the approach into a self-correcting procedure over two dual zero-shot directions, so that incrementally improved translations mutually reinforce each direction.
3 Neural Machine Translation
State-of-the-art NMT systems comprise an encoder, a decoder, and an attention mechanism, which are jointly trained with maximum likelihood in an end-to-end fashion . Among the different variants, two popular ones are the recurrent NMT  and the transformer NMT  models. In both the approaches, the encoder is purposed to map a source sentence into a sequence of state vectors, whereas the decoder uses the previous decoder states, its last output, and the attention model state to infer the next target word (see Figure 2). In a broad sense, the attention mechanism selects and combines the encoder states that are most relevant to infer the next word . In our multi-lingual setting, the decoding process is triggered by specifying the target language identifier (Italian, <it>, in the example of Figure 2). In the following two sub-sections, we briefly summarize the main features of the two considered architectures.
3.1 Recurrent NMT
Recurrent NMT models employ recurrent neural networks (RNNs) to build the internal representations of both the encoder and decoder. Recurrent layers are in general implemented with LSTM  or GRU  units, which include gates able to control the propagation of information over time. While the encoder typically uses a bi-directional RNN, so that both left-to-right and right-to-left word dependencies are captured (see left-hand of Figure 3), the decoder by design can only learn left-to-right dependencies. In general, deep recurrent NMT is achieved by stacking multiple recurrent layers inside both the encoder and the decoder.
While RNNs are in theory the most expressive type of neural networks , they are in practice hard and slow to train. In particular, the combination of two levels of deepness, horizontal along time and vertical across the layers, makes gradient-based optimization of deep RNNs particularly slow to converge and difficult to parallelize . Recent work succeeded in speeding up training convergence  of recurrent NMT by reducing the network size via parameter tying and layer normalization. On the other hand, the simple recurrent NMT model proposed by , which weakens the network time dependencies, has shown to outperform LSTM-based NMT both in training speed and performance.
3.2 Transformer NMT
The transformer architecture  works by relying on a self-attention mechanism, removing all the recurrent operations that are found in the RNN case . In other words, the attention mechanism is re-purposed to also compute the latent space representation of both the encoder and the decoder. The right-hand side of Figure 3 depicts a simple one-layer encoder based on self-attention. Notice that, in absence of recurrence, a positional-encoding is added to the input and output embeddings. Similarly, as the time-step in RNN, the positional information provides the transformer network with the order of input and output sequences. In our work we use absolute positional encoding but, very recently, the use of relative positional information has been shown to improve the network performance .
Overall, the transformer is organized as a stack of encoder-decoder networks that works in an auto-regressive way, using the previously generated symbol as input for the next prediction. Both the decoder and encoder can be composed of uniform layers, each built of sub-layers, i.e., a multi-head self-attention sub-layer and a position-wise feed-forward network sub-layer. Specifically for the decoder, an extra multi-head attentional layer is added to attend to the output states of the encoder. Multi-head attention layers enable the use of multiple attention functions with a computational cost similar to utilizing a single attention.
4 Zero-Shot Self-Training in Multilingual NMT
In this setting, our goal is to improve translation in the zero-shot directions of a baseline multilingual model trained on data covering n languages but not all their possible combinations (see Figure 1). After training a baseline multilingual model with the target-forcing method , our self-learning approach works in the following way:
First, a dual zero-shot inference (i.e., sourcetarget directions) is performed utilizing monolingual data extracted from the training corpus;
Second, the training resumes combining the inference output and the original multilingual data from the non zero-shot directions;
Third, this cycle of training-inference-training is repeated until a convergence point is reached on the dual zero-shot directions.
Notice that, at each iteration, the original training data is augmented only with the last batch of generated translations. We observe that the generated outputs initially contain a mix of words from the shared vocabulary but, after few iterations, they tend to only contain words in the zero-shot target language thus becoming more and more suitable for learning. The training and inference strategy of the proposed approach is summarized in Algorithm 1, whereas the flow chart (see Figure 4) further illustrates the training and inference pipeline.
|Algorithm 1: Train-Infer-Train (TIF)|
|1: TIF: (, , )|
|2: Train(, ) //train multilingual base model on data|
|3: Extract(, ) //extract monolingual data from|
|4: Extract(, ) //extract monolingual data from|
|5: for , do|
|6: Infer(, , ) //translate into|
|7: Infer(, , ) //translate into|
|8: //augment original data|
|9: Train(, ) //re-train model on augmented data|
|10: end for|
The proposed approach is performed in three steps, where the latter two are iterated for a few rounds. In the first step (line 2), a multilingual NMT system is trained from scratch on the available data (”Train” step). In the second step (lines 7-8), the last trained model is run to translate (”Infer” step) between the zero-shot directions monolingual data and extracted from (lines 3-4). Then, in the third step (line 10), training of is re-started on the original data plus the generated synthetic translations and , by keeping the extracted monolingual data and always on the target side (”Train” step). The updated model is then again used to generate synthetic translations, on which to re-train , and so on.
In the multilingual NMT scenario, the automatic translations used as the source part of the extended training data will likely contain a mixed-language that includes words from a vocabulary shared with other languages. The expectation is that, round after round, the model will generate better outputs by learning at the same time to translate and “correct” its own translations by removing spurious elements from other languages. If this intuition holds, the iterative improvement will yield increasingly better results in translating between the source target zero-shot directions.
5.1 NMT Settings
We trained and evaluated multilingual NMT systems based on the RNN  and transformer  models. Table 2 summarizes the hyper-parameters used for all our models. The RNN experiments are carried out using the NMT toolkit Nematus222https://github.com/EdinburghNLP/nematus , whereas the transformer models are trained using the open source OpenNMT-tf333https://github.com/OpenNMT/OpenNMT-tf toolkit .
Training and inference hyper-parameters for both approaches and toolkits are fixed as follows. For the RNN experiments, the Adagrad  optimization algorithm is utilized with an initial learning rate of 0.01 and mini-batches of size 100. Considering the high data sparsity of our low-resource setting and to prevent model over-fitting , we applied a dropout on every layer, with probability on the embeddings and the hidden layers, and on the input and output layers. For the experiments using the transformer approach, a dropout of is used globally. To train the baseline multilingual NMT, we use Adam  as the optimization algorithm with an initial learning rate scale constant of . For the transformer, the learning rate is increased linearly in the early stages (warmup_training_steps=); after that, it is decreased with an inverse square root of training step . In all the reported experiments, the baseline models are trained until convergence, while each training round after the inference stage is assumed to iterate for 3 epochs. In case of the transformer NMT, M4-NMT (four translation directions multilingual system), and M6-NMT (six translation directions multilingual system) BLEU scores are computed using averaged model from the last seven checkpoints in the same training run . For decoding, a beam search of size 10 is applied for recurrent models, whereas one of size is used for transformer models.
5.2 Dataset and preprocessing
We run all our experiments on the multilingual translation shared task data released for the 2017 International Workshop on Spoken Language Translation (IWSLT)444http://workshop2017.iwslt.org/. In particular, we used the subset of training data covering all possible language pair combinations between Italian, Romanian, and English . For development and evaluation of the models, we used the corresponding sets from the IWSLT2010  and IWSLT2017 evaluation campaigns. Details about the used data sets are reported in Table 3. At the preprocessing stage, we applied word segmentation for each training condition (i.e. bilingual or multi-lingual) by learning a sub-word dictionary via Byte-Pair Encoding , by setting the number of merging rules to 39,500. We observe a rather high overlap between the language pairs (i.e the English dataset paired with Romanian is highly similar to the English paired with Italian). Because of this overlapping, the actual unique sentences in the dataset are approximately half of the total size. Consequently, on one side, this exacerbates the low-resource aspect in the multilingual models while, on the other side, we expect some positive effect on the zero-shot condition. The final size of the vocabulary, both in case of the bilingual and the multilingual models, stays under 40,000 sub-words. An evaluation script to compute the BLEU  score is used to validate models on the dev set and later to choose the best performing models. Furthermore, significance tests computed for the BLEU scores are reported using Multeval .
We trained models for two different scenarios. The first one is the multi-lingual scenario with all the available language pairs, while the second one is for the zero-shot and pivoting approaches which excludes parallel sentences from the training data. For both scenarios, we have also trained bilingual RNN and Transformer models for comparing bilingual against multilingual systems and for comparing pivoting with bilingual and multilingual models.
6 Models and Results
6.1 Bilingual Vs. Multilingual NMT
We compare the translation performance of six independently-trained bilingual models against one single multilingual model trained on the concatenation of all the six language pairs datasets, after prepending the language flag on the source side of each sentence. The performance of both types of systems is evaluated on test2017 and reported in Table 4. The experiments show that a multilingual system outperforms the bilingual systems with variable margins. The improvements, which are observed in all the language directions, are likely brought by the cross-lingual parameter transfer between the additional language pairs involved in the source and target side.
Table 4 shows that the transformer model is definitely superior to the RNN model for all directions and set-ups. With a larger margin of +3.22 (NMT) and +4.21 (M6-NMT), the transformer outperforms the RNN in the It-En direction. The closest performance between the two approaches is observed in the Ro-It direction, with the transformer showing a +1.37 (NMT) and +1.09 (M6-NMT) BLEU score increase compared to the RNN counterpart. Moreover, multilingual architectures in general outperform their equivalent models trained on single language pairs. The highest improvement of the M6-NMT over the NMT systems is observed when the target language is English. For instance, in the It-En direction, the multilingual approach gained +1.94 (RNN) and +2.93 (Transformer) over the single language pair models. Similarly, a +1.80 (RNN) and +1.85 (Transformer) gains are observed in the Ro-En direction. However, the smallest gain of the multilingual models occurred when translating into either Italian or Romanian. Independently from the target language in the experimental setting, the slight difference in the dataset size (that tends to benefit the English target, see Table 3) showed to impact the performance on non-English target directions.
6.2 Pivoting using a Multilingual Model
The pivoting experiment is setup by dropping the ItalianRomanian parallel segments from the training data, and by training i) a new multilingual-model covering four directions and ii) a single model for each language direction (It En, En It, Ro En, En Ro). Our main aim is to analyze how a multilingual model can improve a zero-shot translation task using a pivoting mechanism with English as a bridge language in the experiment. Moreover, the use of a multilingual model for pivoting is motivated by the results we acquired using the M6-NMT (see Table 4).
|It En Ro||16.3||17.58||+1.28||16.59||16.77||+0.18|
|Ro En It||18.69||18.66||-0.03||17.87||19.39||+1.52|
The results in Table 5 show the potential, although partial, of using multilingual models for pivoting unseen translation directions. The comparable results achieved in both directions speak in favor of training and deploying one system instead of two distinct NMT systems. Remarkably, the marked difference between RNN and transformer is vanished in this condition. Pivoting using the M4-NMT system showed to perform better in three out of four evaluations, from the RNN and transformer runs. Note that the performance of the final translation (i.e pivot-target) is subject to the noise that has been propagated from the source-pivot translation step. Meaning pivoting is a favorable strategy when we have strong models in both directions of the pivot language.
6.3 Zero-shot Translations
For the direct zero-shot experiments and the application of the train-infer-train strategy, we only carried out experiments with the transformer approach. Preliminary results showed its superiority over the RNN together with the possibility to carry out experiments faster and with multiple GPUs.
In this experiment, we show how our approach helps to significantly boost the baseline multilingual NMT model. We run the train-infer-train for five consecutive stages, where each round consists in 2-3 epochs of additional training on the augmented training data. Table 6 shows the improvements on the dual ItalianRomanian zero-shot directions.
In both zero-shot directions the gain in a larger margin comes using the M-NMT model at R1. This is the first model trained after the inclusion of the dataset generated by the dual-inference stage. The ItRo direction improves by BLEU points from a to , whereas RoIt improves from a baseline score of to BLEU (). The contribution of the self correcting process can be seen in the subsequent rounds, i.e., the improvements after each inference stage suggest that the generated data are getting cleaner and cleaner. With respect to the Transformer model pivoting results shown in Table 5, our approach outperformed both single pair and multilingual pivoting methods at the second round (R2) (see the third column of Table 6). Compared with the better performing multilingual pivoting, our approach at the fifth round (R5) has a and BLEU gain for the ItRo and RoIt directions respectively.
In addition to outperforming the pivoting mechanism, an interesting trend arises when we compare our approach with the results of the single language pair and multilingual models reported in Table 4. The summary in Table 7 shows the effectiveness of a dual-inference mechanism in allowing the model to learn from its outputs. Compared to the models trained using parallel data (i.e., NMT and M6-NMT), our approach (R5) is either comparable ( BLEU in ItRo) or better performing ( BLEU in RoIt). The trend from the train-infer-train stages indicates that, with additional rounds, our approach can further improve the dual translations. Overall, our iterative self-learning approach showed to deliver better results than the bilingual counterparts within five rounds, where each rounds iterates for a maximum of three epochs. Indeed, the improvement from our approach is a concrete example to train models in a self-learning way, potentially benefiting language directions with a parallel data, if casted in a similar setting.
In this paper, we used a multilingual NMT model in a low-resource language pairs scenario. Integrating and extending the work presented in  and , we showed that a single multilingual system outperforms bilingual baselines while avoiding the need to train several single language pair models. In particular, we confirmed the superiority of transformer over recurrent NMT architectures in a multilingual setting. For enabling and improving a zero-shot translation, we showed i) how a multilingual pivoting can be used for achieving comparable results to those of multiple bilingual models, and ii) that our proposed self-learning procedure boosts performance of multilingual zero-shot directions by even outperforming both pivoting and bilingual models. In future work, we plan to explore our approach across language varieties using a multilingual model.
This work has been partially supported by the EC-funded projects ModernMT (H2020 grant agreement no. 645487) and QT21 (H2020 grant agreement no. 645452). This work was also supported by The Alan Turing Institute under the EPSRC grant EP/N510129/1 and by a donation of Azure credits by Microsoft. We gratefully acknowledge the support of NVIDIA Corporation with the donation of the Titan Xp GPU used for this research.
-  (2014) Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473. Cited by: §1, §3.
-  (2018) Neural versus phrase-based mt quality: an in-depth analysis on english-german and english-french. Computer Speech & Language 49, pp. 52–70. Cited by: §1.
-  (2009) Domain adaptation for statistical machine translation with monolingual resources. In Proceedings of the fourth workshop on statistical machine translation, pp. 182–189. Cited by: §2.2.
-  (2016) Findings of the 2016 conference on machine translation (wmt16). In Proceedings of the First Conference on Machine Translation (WMT), Vol. 2, pp. 131–198. Cited by: §1.
-  (2011) Bootstrapping arabic-italian smt through comparable texts and pivot translation. In 15th Annual Conference of the European Association for Machine Translation (EAMT), Cited by: §2.2.
-  (28-30) WIT: web inventory of transcribed and translated talks. In Proceedings of the 16 Conference of the European Association for Machine Translation (EAMT), Trento, Italy, pp. 261–268. Cited by: §1, §5.2.
-  (2016) The iwslt 2016 evaluation campaign. Proc. of IWSLT, Seattle, pp. 14, WA, 2016.. Cited by: §1.
-  (2014) On the properties of neural machine translation: encoder-decoder approaches. arXiv preprint arXiv:1409.1259. Cited by: §2.1, §3.1.
-  (2014) Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078. Cited by: §5.1.
-  (2011) Better hypothesis testing for statistical machine translation: controlling for optimizer instability. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies: short papers-Volume 2, pp. 176–181. Cited by: §5.2.
-  (2018-05) Deep neural machine translation with weakly-recurrent units. In Proceedings of the 21st Annual Conference of the European Association for Machine Translation (EAMT), Alicante, Spain. Cited by: §3.1.
-  (2011) Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research 12 (Jul), pp. 2121–2159. Cited by: §5.1.
-  (2016) Zero-resource translation with multi-lingual neural machine translation. arXiv preprint arXiv:1606.04164. Cited by: §2.2.
-  (2006) Scalable inference and training of context-rich syntactic translation models. In Proceedings of the 21st International Conference on Computational Linguistics and the 44th annual meeting of the Association for Computational Linguistics, pp. 961–968. Cited by: §1.
-  (2016) Toward multilingual neural machine translation with universal encoder and decoder. arXiv preprint arXiv:1611.04798. Cited by: §1, §1, §2.2, §2.2.
-  (1997) Long short-term memory. Neural computation 9 (8), pp. 1735–1780. Cited by: §3.1.
-  (2016) Google’s multilingual neural machine translation system: enabling zero-shot translation. arXiv preprint arXiv:1611.04558. Cited by: §1, §1, §2.2, §2.2, §4.
-  (2016) The amu-uedin submission to the wmt16 news translation task: attention-based nmt models as feature functions in phrase-based smt. arXiv preprint arXiv:1605.04809. Cited by: §5.1.
-  (2014) Adam: a method for stochastic optimization. arXiv preprint arXiv:1412.6980. Cited by: §5.1.
-  (2017) OpenNMT: open-source toolkit for neural machine translation. arXiv preprint arXiv:1701.02810. Cited by: §5.1.
-  (2017) Six challenges for neural machine translation. arXiv preprint arXiv:1706.03872. Cited by: §1.
-  (2017) Multilingual neural machine translation for low resource languages. In CLiC-it 2017 – 4th Italian Conference on Computational Linguistics, to appear, Cited by: §2.2, §7, footnote 1.
-  (2017) Improving zero-shot translation of low-resource languages. In 14th International Workshop on Spoken Language Translation, Cited by: §7, footnote 1.
-  (2015) Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025. Cited by: §3.
-  (2002) BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting on association for computational linguistics, pp. 311–318. Cited by: §5.2.
-  (2010) Overview of the iwslt 2010 evaluation campaign. In International Workshop on Spoken Language Translation (IWSLT) 2010, Cited by: §5.2.
-  (2017-09) The university of edinburgh’s neural mt systems for wmt17. In Proceedings of the Second Conference on Machine Translation, Volume 2: Shared Task Papers, Copenhagen, Denmark, pp. 389–399. External Links: Cited by: §3.1.
-  (2017) Nematus: a toolkit for neural machine translation. arXiv preprint arXiv:1703.04357. Cited by: §5.1.
-  (2015) Improving neural machine translation models with monolingual data. arXiv preprint arXiv:1511.06709. Cited by: §2.2.
-  (2015) Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909. Cited by: §5.2.
-  (2018) Self-attention with relative position representations. arXiv preprint arXiv:1803.02155. Cited by: §3.2.
-  (1995) On the computational power of neural nets. Journal of Computer and System Sciences 50 (1), pp. 132 – 150. External Links: Cited by: §3.1.
-  (2014) Dropout: a simple way to prevent neural networks from overfitting.. Journal of machine learning research 15 (1), pp. 1929–1958. Cited by: §5.1.
-  (2014) Sequence to sequence learning with neural networks. In Advances in neural information processing systems, pp. 3104–3112. Cited by: §3.
-  (1989-06) Language transfer-cross-linguistic influence in language learning. Cambridge University Press. Cambridge Books Online., pp. 222. Cited by: §1.
-  (2017) Attention is all you need. In Advances in Neural Information Processing Systems, pp. 6000–6010. Cited by: §3.2, §3, §5.1, §5.1.
-  (2016) Google’s neural machine translation system: bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144. Cited by: §3.1.
-  (2016) Dual learning for machine translation. CoRR abs/1611.00179. Cited by: §2.2.