EMPLOYING PIVOT LANGUAGE TECHNIQUE THROUGH STATISTICAL AND NEURAL MACHINE TRANSLATION FRAMEWORKS: THE CASE OF UNDER-RESOURCED PERSIAN-SPANISH LANGUAGE PAIR

International Journal on Natural Language Computing (IJNLC) Vol. 6, No.5, October 2017
DOI: 10.5121/ijnlc.2017.6503 37
EMPLOYING PIVOT LANGUAGE TECHNIQUE
THROUGH STATISTICAL AND NEURAL MACHINE
TRANSLATION FRAMEWORKS: THE CASE OF
UNDER-RESOURCED PERSIAN-SPANISH LANGUAGE
PAIR
Benyamin Ahmadnia and Javier Serrano
Autonomous University of Barcelona, Cerdanyola del Valles, Spain
ABSTRACT
The quality of Neural Machine Translation (NMT) systems like Statistical Machine Translation (SMT)
systems, heavily depends on the size of training data set, while for some pairs of languages, high-quality
parallel data are poor resources. In order to respond to this low-resourced training data bottleneck reality,
we employ the pivoting approach in both neural MT and statistical MT frameworks. During our
experiments on the Persian-Spanish, taken as an under-resourced translation task, we discovered that, the
aforementioned method, in both frameworks, significantly improves the translation quality in comparison
to the standard direct translation approach.
KEYWORDS
Statistical Machine Translation, Neural Machine Translation, Pivot Language Technique
1. INTRODUCTION
The purpose of the statistical machine translation is to translate a source language sequences into
a target language ones by assessing the plausibility of the source and the target sequences in
relation to existing bodies of translation between the two languages. A huge shortcoming in SMT
is the lack of consistent parallel data for many language pairs and corpora of this type [2]. In order
to overcome this shortcoming, researchers have developed different ways to connect source and
target languages with only a small parallel corpus, that is used to generate a systematic SMT when
a proper bilingual corpus is lacking or the existing ones are weak [5, 10, 13, 28, 29]. This is an
important issue when there are languages with inefficient NLP (Natural Language Processing)
resources that are not able to provide an SMT system. Nevertheless, there are sufficient resources
between them and some other languages.
Afterwards, the goal of neural machine translation is to build a single neural network that can be
jointly tuned to maximize the translation quality [26]. The NMT has built state-of-the-art for many
pairs of languages only by using parallel training data set, and has shown competitive results in
recent researches [3, 20, 26]. In comparison with conventional SMT [22], competitive translation
quality has been obtained on well-resourced pairs of languages such as English-French or
German-English.
In spite of these achievements, there are also some shortcomings. The NMT systems indicate
poorer performance in comparison to a standard tree-to-string SMT system for under-resourced
pairs of languages, because the neural network is a data-driven approach [31]. The NMT is non-
trivial because it directly maximizes the probability of the target sentences given the source

38
sentences without modeling latent structures. In order to bridge the source-target translation
model through the source-pivot and the pivot-target translation models, we need to use a joint
training for the NMT. We have to make the translation path from the source language to the target
one, with the bridge language translations. We investigate a kind of connection terms, which uses
a small source-target parallel corpus to guide the translation path with the bridge language
translations, so that we can connect these two directional models.
The remainder of this article is organized as follows; We introduce the structures of both
translation frameworks in Section 2, and the concepts of the pivoting method in Section 3. In
Section 4 we describe and analyze the experiments. Section 5 describes the related works, and
Section 6 gives a conclusion of the article.
2. TRANSLATION FRAMEWORKS
In this section we will introduce the architectures of the statistical machine translation systems,
and the neural machine translation systems, which are used to deal with our experiments and the
translation process.
2.1. Statistical MT Framework
The statistical machine translation paradigm has, as its most important elements, the idea; that
probabilities of the source and the target sentences can find the best translations. Frequently used
paradigms of SMT on the log-linear model are the phrase-based, the hierarchical phrase-based,
and the ngram-based. In our experiments we use the phrase-based SMT system with the
maximum entropy framework [4]:
The phrase-based SMT model is an example of the noisy-channel approach, where we can present
the translation hypothesis (t) as the target sentence (given (s) as a source sentence), maximizing a
log-linear combination of feature functions:
This equation called the log-linear model, where λm corresponds to the weighting coefficients of
the log-linear combination, and the feature functions hm(s,t) to a logarithmic scaling of the
probabilities of each model. The translation process involves segmenting the source sentences
into source phrases, translating each source phrase into a target phrase, and reordering these target
phrases to yield the target sentence.
2.2. Neural MT framework
Neural machine translation aims at designing a comprehensible trainable model. In this model, all
components are tuned based on a training corpora to raise the translation accuracy and
performance. Building and training a single, large neural network that reads a sentence and
outputs a correct translation are the chief purposes of NMT. Any neural network which maps a

39
source sentence to a target one is considered as an NMT system, where all sentences are assumed
to terminate with a special “end-of-sentence” (<eos>) token. More concretely, an NMT system
uses a neural network to parameterize the following conditional distributions for 1 ≤ j ≤ m:
By doing so, it becomes possible to compute and therefore maximize the log probability of the
target sentence given the source sentence:
There are many ways to parameterize these conditional distributions. For example, Kalchbrenner
and Blunsom (2013) used a combination of a convolutional neural network and a Recurrent
Neural Network (RNN) [20], Sutskever et al. (2014) used a deep Long/Short-Term Memory
(LSTM) model [26], Cho et al. (2014) used an architecture similar to the LSTM [8], and Bahdanau
et al. (2015) used a more elaborate neural network architecture that uses an attentional mechanism
over the input sequence [3, 15].
3. PIVOT LANGUAGE TECHNIQUE
Translation systems in terms of both SMT and NMT, have made great strides in translation
quality. State-of-the-art have shown that, high-quality translation output is dependent on the
availability of massive amounts of parallel texts in the source and the target languages. However,
there are a large number of languages that are considered low-density, either because the
population speaking those languages is not very large, or even if millions of people speak those
languages, insufficient amounts of parallel texts are available in those languages.
This technique is an idea to generate a systematic machine translation when a proper bilingual
corpus is lacking or the existing ones are weak. This article shows that, how such corpora can be
used to achieve high translation quality through the pivot language technique, and we investigate
the performance of this strategy through our considered translation frameworks.
3.1. Pivoting Strategy for SMT
According to [29], pivot-based strategies that employed for SMT systems can be classified into
these categories:
1. The “transfer method” also known as cascade or sentence translation pivot strategy, which
translates the text in the source language to the pivot, using a source-pivot translation model, and
then to the target language using a pivot-target translation model.
2. The “multiplication method” also identified as triangulation or phrase translation pivot
strategy, which merges the corresponding translation probabilities of the translation models for
the source-pivot and the pivot-target languages, generates a new source-target translation model.
3. The “synthetic corpus method” which tries to create a synthetic source-target corpus by
translating the pivot part in the source-pivot corpus, into the target language with a pivot-target
model, and translating the pivot part in the target-pivot corpus, into the source language with a

40
pivot-source model. Finally combining the source sentences with the translated target sentences
or combining the target sentences with the translated source sentences. Nevertheless, it is
somehow difficult to build a high-quality translation system with a corpus created only by a
machine translation system.
In this article our SMT pivoting experiments just rely on the first and the second methods.
3.1.1. Transfer Method
In the sentence translation pivot strategy, we first translate the Persian sentences into the English
ones, and then translate these English sentences into the Spanish ones separately. We select the
highest scoring sentence from the Spanish sentences.
In this technique for assigning the best Spanish candidate sentence (s) to input the Persian
sentence (p), we maximize the probability P(s|p) by defining hidden variable (e), which stands
for the pivot language sentences, we gain:
In Equation (6), summation on all (e) sentences is difficult, so we replace it by maximization, and
Equation (7) is an estimate of Equation (6):
Instead of searching all the space of (e) sentences, we can just search a subspace of it. For
simplicity we limit the search space in Equation (8). A good choice is (e) subspace produced by
the (n-best) list output of the first SMT system (source-pivot):
In fact each sentence (p) of the Persian test set is mapped to a subspace of total (e) space and
search is done in this subspace for the best candidate sentence (s) of the second SMT system
(pivot-target).
3.1.2. Multiplication Method
For applying the phrase translation pivot strategy, we directly construct the Persian-Spanish
phrase translation table from the Persian-English, and the English-Spanish phrase-tables.
In this technique phrase (p) in the source-pivot phrase-table is connected to (e), and phrase (e) is
associated with (s) in the pivot-target phrase-table. We link (p) and (s) in the new phrase-table for
the source-target. For scoring the pair phrases of the new phrase-table, assuming P(e|p) as the
score of the Persian-English phrases and P(s|e) as the score of the English-Spanish phrases, then

41
the score of the new pair phrases (p) and (s), P(s|p), in the Persian-Spanish phrase-table is:
(e) is a hidden variable and actually stands for the phrases of pivot language:
If we assume that, (p) and (s), are independent given (e):
For simplicity the summation on all the (e) phrases is replaced by maximization, then Equation
(11) is approximated by:
3.2. Pivoting Strategy for NMT
Considering P(p|s; θsp) and P(t|p; θpt) as the source-pivot and the pivot-target NMT models
respectively, while giving two parallel corpora, the source-pivot parallel corpus (Csp) and the
pivot-target parallel corpus (Cpt). We employ the pivoing strategy in which the target sentence is
generated for a source sentence after it is first translated to the pivot sentences. The crucial point
is to jointly instruct two translation models, P(p|s; θsp) and P(t|p; θpt), heading at establishing the
source-target translation path with the pivot sentences as the intermediate translations:
The source-pivot Likelihood, the pivot-target Likelihood, and the linking term, are the main
objectives of our training model. In order to balance the significance between the Likelihoods and
the linking term, (λ) is utilised. The linking term includes two sets of parameters; (θsp) and (θpt),
for the source-pivot and the pivot-target translation models respectively. The linking term is
controlled so as to allow two independently trained parameters from two different translation
models to interact mutually. Replacing the linking term by any function with the parameters of
these two included directional NMT models is feasible.
In general, for many language pairs and domains, small corpora are pervasive. Given a test source
sentence, it will be translated to the target sentence eventually through the pivoting technique.
This translation path will be reinforced with the supply for parallel sentence pairs between the
source and the target. The employed approach in the current study treats the pivot sentences as
latent variables:

42
Where (p) is a latent pivot sentence. The intuition of Equation (14) is to maximize the translation
probability of the target sentences given the source sentences via the pivot candidate translations.
The source-pivot translation model first transforms the source sentences into the latent pivot
sentences, from which, the pivot-target translation model aims to construct the target sentences.
This training criterion conforms to the pivot translation strategy adopted by the test procedure [6].
The partial derivative of J(θsp, θpt) with respect to the parameters (θsp) of the source-pivot model is
calculated as:
The partial derivative with respect to the parameters (θpt) is similar to Equation (15). In our
connection term, if we continue to expand the last term of Equation (15), a challenge emerges:
Enumerating all of pivot candidate translations p ∈P(s) in Equation (16) is intractable because of
the exponential search space for the pivot translations. As an alternative solution, the subset
approximation is normally employed. In order to approximate the full space, we utilized a subset
⊂P΄(s) P(s). In addition, we undertook two methods to generate P΄(s), sampling (k) translations
from the full space and generating (k-best) list of candidate translations. The findings revealed
that generating (k-best) list operates better.
Holding three parallel corpora including the source-pivot, the pivot-target, and the source-target,
we still utilize mini-batch stochastic gradient descent algorithm in order to update the parameters.
Though three mini-batches of parallel sentence pairs are randomly picked in each iteration from
the source-pivot, the pivot-target and the source-target parallel corpora. Likelihood, in order to
get the (k-best) pivot translations, decoding the source sentences of the source-target mini-batch is
needed. Afterwards, the gradients for these batches are calculated and then collected for
parameter updating purposes. The decision rules for the source-pivot and the pivot-target NMT
models are respectively given by:

43
4. EXPERIMENTAL FRAMEWORK
In this section, we present a set of experiments on both SMT and NMT frameworks including the
pivot language technique to overcome the limitation of training resources scarsity. Then we
present our results to compare the Persian-Spanish translation quality in either both
aforementioned frameworks.
Our data resources in both SMT and NMT experiments are collected from the in-domain Tanzil
parallel corpus [27]. In this corpus the Persian-Spanish part contains more than (68K) sentences
and approximately (3.51M) words, the Persian-English part contains more than (1M) sentences
and more than (57M) words, and the English-Spanish part contains more than (133K) sentences
and approximately (4.25M) words. Table 1 shows our data resource statistics.
Table 1. Corpus Statistics.
Corpus Direction Sentences
Tanzil Persian - English 1,028,996
Tanzil English - Spanish 133,735
Tanzil Persian - Spanish 68,601
The training part of our system involved of (60K) sentences. For the tuning and the testing steps
we collected parallel texts from the Tanzil corpus, we extracted (3K) sentences for the tuning, and
(5K) sentences for the testing.
4.1. SMT Systems Experiments and Results
“MOSES” package [21], is used for training our SMT systems. Through utilising MOSES decoder,
we apply fast-align approach [12], for sentence alignment in our experiment. The employed
language model for all SMT systems are 3-grams and they are built using the KenLM toolkit [19].
We use the BLEU metric [24], in order to evaluate the systems performance. Table 2 presents the
results of the Persian-English, the English-Spanish, and the Persian-Spanish direct translation
systems.
Table 2. The BLEU scores of the Pe-En, the En-Es, and the Pe-Es direct SMT systems.
System Persian - English English - Spanish Persian - Spanish
Direct 14.31 15.34 11.39
In the other portion of this experiment the two phrase-tables employed to shape a new table in the
phrase pivoting method are extracted in turn from the Persian-English and the English-Spanish
translation systems. Table 3 illustrates the results of the sentence translation pivoting and the
phrase translation pivoting of the Persian-Spanish translation system through English as the
intermediary language.

44
Table 3. The BLEU scores of the Pe-(En)-Es pivoting SMT systems.
System Phrase - Level Sentence - Level
Pivoting 13.55 12.78
According to the results, in the case of Persian-Spanish language pair, the pivot-based translation
method is suitable for the scenario that there exist large amounts of source-pivot and pivot-target
bilingual corpora and only a little source-target bilingual data. Thus we selected (60K) sentence
pairs from the source-target bilingual corpora to simulate the lack of source-target data.
4.2. NMT Systems Experiments and Results
“MANTIS” package [9], is used as the attention-based NMT systems in our experiments. We have
tried to analyze the Persian-Spanish language pair through English as the bridge language. For
this language pair, we removed the empty lines and retain sentence pairs with no more than (50)
words. In order to avoid the constitution of the tri-lingual corpora by the source-pivot and the
pivot-target, the overlapping section of the pivot sentences from the source-pivot and the pivot-
target corpora should be divided into two equal parts and also they should be combined separately
with the non-overlapping parts. For the language modeling we used the RNN language model
[23], separately. In order to use the Likelihood linking term, we set the sample size, (k), to (40), in
order to avoid the weird segmentation fault error message. The hyper-parameter, (λ), to (1.0), and
the threshold of gradient clipping to (0.1). The parameters for the source-pivot and the pivot-
target translation models in Likelihood are initialized by pre-trained model parameters.
All the sentences of the corpora are encrypted by the tokenize.perl script and the development and
the test data sets are from the Tanzil corpus as well as the training data set. The evaluation metric
is BLEU [24], as calculated by the multi-bleu.perl script. We have used English as the pivot
language and followed Likelihood linking term that jointly train the source-pivot and the pivot-
target translation models. We have tried to show a comparison between translation quality for the
source-pivot, the pivot-target, and the source-target directions. The source-target translation
results are obtained by translating pivot sentences. Table 4 shows a comparison results on the
Persian-Spanish translation task from the Tanzil corpus.
Table 4. The BLEU scores comparing the direct with the Likelihood NMT system.
System Persian - English English - Spanish Persian - Spanish
Direct 14.17 14.88 11.19
Likelihood 14.31 15.02 12.93
The results show that, the BLEU scores of the Likelihood method are better than the standard
direct training. Our analysis points out that, the Likelihood strategy improves the translation
performance on the Persian-Spanish translation task up to (1.74) BLEU scores (in comparison
with the direct translation approach), by introducing the source-target parallel corpus to maximize
P(t|s; θsp, θpt) with (p) as the latent variables makes the source-pivot and the pivot-target
translation models improved collaboratively. As we have showed, this approach improves
translation quality of both pivot and target sentences.
5. RELATED WORKS
In the case of low-resourced language pairs, some researchers introduce a pivot language to
bridge source and target languages in SMT, such as the case of Catalan-English with no parallel

45
corpus [11]. Some researchers investigated the SMT system with pivot language method. One
example is Hartley et al. (2007) who used the Russian language as a pivot for translating from
Ukrainian to English. From their experience, we figured out that, it is possible to achieve better
translation quality with pivot language approach [18]. Habash and Hu (2007) compared two
approaches for Arabic-Chinese language pair with direct MT system through English as a pivot
language. Their researches indicate that using English as a pivot language in either approaches
leads to better results than the direct translation from Arabic to Chinese [17]. Going in the same
direction, Al-Hunayti et al. (2010) presented a comparison between two common pivot strategies;
phrase translation and sentence translation, in order to enhance Danish-Arabic SMT system. This
approach showed that the sentence pivoting overtakes the phrase pivoting when common parallel
corpora are not available [1].
Firat et al. (2016) proposes a multiway, multilingual NMT model that enables zero-resourced MT.
In order to find tune parameters of the low-resourced language pairs using trained parameters on
the high-resourced language pairs [14]. Zoph et al. (2016) adopted a transfer learning method. The
aim was to build a source-target NMT model. Because of limited quantity, quality, and coverage
for parallel corpora, additional data resource have come under scrutiny lately [31]. For example,
Zhang and Zong (2016) proposed two approaches to incorporate the source side monolingual
corpora; One is to employ self-training algorithm to generate parallel corpora from monolingual
corpora. The other adopts multi-task learning framework to enhance the encoder network of NMT
[30]. On the other hand, Cheng et al. (2016) introduced an auto-encoder framework to reconstruct
monolingual sentences using the source-target and the target-source NMT models [7]. Researchers
such as Gulccehre et al. (2015) proposed to incorporate the target side monolingual corpora as the
language model for NMT [16]. As Sennrich et al. (2016) pairs the target monolingual corpora with
its corresponding translations then merges them with parallel data for retraining the source-target
model [25].
6. CONCLUSION
In this article, we have tried to analyze the behavior of the pivot (bridge) language technique on
both statistical and neural machine translation systems for the Persian-Spanish, which is a
resource poor language pair.
In the first case, we have compared two common pivoting translation methods comprising the
phrase-level combination, and the sentence-level combination, for the Persian-Spanish SMT by
employing English as an intermediary language. By organizing controlled experiments, we have
assessed the performances of these two methods against the performance of directly trained SMT
system. The results revealed that utilizing English as a bridging language in either approaches
gives better results than by the direct translation approach from Persian to Spanish.
In the second case, we have presented a joint training method for the Persian-Spanish NMT via
English as a bridge language. The connection term in our joint training objective makes the
Persian-English and the English-Spanish translation models interact better. So that the
experiments confirm that, this approach achieves significant improvements.
ACKNOWLEDGEMENTS
We would like to express our sincere gratitude to Shekoofeh Dadgostar for all her support. We
have benefited from her erudition and thoughtful comments which truly enriched this article.

46
REFERENCES
[1] Al-Hunaity, Mossab & Maegaard, Bente & Hansen, Dorte, (2010) “Using English as a Pivot
Language to Enhance Danish-Arabic Statistical Machine Translation”.
[2] Babych, Bogdan & Hartley, Anthony & Sharoff, Serge & Mudraya, Olga, (2007) “Assisting
Translators in Indirect LexicalTransfer”, ACL.
[3] Bahdanau, Dzmitry & Cho, Kyunghyun & Bengio, Yoshua, (2015) “Neural machine translation by
jointly learning to align and translate”, ICLR.
[4] Berger, Adam.L & Della Pietra, Stephen & Della Pietra, Vincent.J, (1996) “A Maximum Entropy
Approach to Natural Language Processing”, Journal of Computational Linguistics.
[5] Bertoldi, Nicola & Barbaiani, Madalina & Federico, Marcello Cattoni, Roldano, (2008) “Phrase-based
statistical machine translationwith pivot languages”, IWSLT.
[6] Cheng, Y.ong & Liu, Yang & Yang, Qian & Sun, Maosong & Xu, Wei, (2016) “Neural Machine
Translation with Pivot Languages”, ACL.
[7] Cheng, Yong & Xu, Wei & He, Zhongjun & He, Wei & Wu, Hua & Sun, Maosong & Liu, Yang,
(2016) “Semi-Supervised Learning for Neural Machine Translation”, ACL.
[8] Cho, Kyunghyun & Merrienboer, Bert.V & Gulccehre, Caglar & Bougares, Fethi & Schwenk, Holger
& Bengio, Yoshua, (2014) “Learning Phrase Representations using RNN Encoder-Decoder for
Statistical Machine Translation”.
[9] Cohn, Trevor & Hoang, Cong.D.V & Vymolova, Ekaterina & Yao, Kaisheng & Dyer, Chris &
Haffari, Gholamreza, (2016) “Incorporating Structural Alignment Biases into an Attentional Neural
Translation Model”, NAACL-HLT.
[10] Cohn, Terevor & Lapata, Mirella, (2007) “Machine Translation by Triangulation: Making Effective
Use of Multi-Parallel Corpora”.
[11] DeGispert, Adria & Marino, Jose.B, (2006) “Catalan-English statistical machine translation without
parallel corpus: bridging through Spanish”.
[12] Dyer, Chris & Chahuneau, Victor & Smith, Noah.A, (2013) “simple, Fast, and Effective
Reparameterization of IBM Model 2”.
[13] El-Kholey, Ahmed & Habash, Nizar, (2013) “Language Independent Connectivity Strength Features
for Phrase Pivot Statistical Machine Translation”.
[14] Firat, Orhan & Sankaran, Baskaran & Al-Onaizan, Yaser & Yarman Vural, Fatos.T & Cho,
Kyughyun, (2016) “Zero-Resource Translation with Multi-Lingual Neural Machine Translation”.
[15] Graves, Alex & Wayne, Greg & Danihelka, Ivo, (2014) “ Neural Turing Machines”.
[16] Gulccehre, Caglar & Firat, Orhan & Xu, Kelvin & Cho, Kyunghyun & Barrault, Loic & Lin, Huei-
Chi & Bougares, Fethi & Schwenk, Holger & Bengio, Yoshua, (2015) “On Using Monolingual
Corpora in Neural Machine Translation”.
[17] Habash, Nizar & Hu, Jun, (2007) “Improving Arabic-Chinese Statistical Machine Translation using
English as Pivot Language”.
[18] Hartley, Anthony & Babych, Bogdan & Sharoff, Serge, (2007) “Translating from under-resourced
languages Comparing direct transfer against pivot translation”.
[19] Heafield, Kenneth & Pouzyrevsky, Ivan & Clark, Jonathan.H & Koehn, Philipp, (2013) “Scalable
Modified Kneser-Ney Language Model Estimation”, ACL.
[20] Kalchbrenner, Nal & Blunsom, Phil, (2013) “Recurrent Continuous Translation Models”.
[21] Koehn, Philipp & Hoang, Hieu & Birch, Alexandra & Callison-Burch, Chris & Federico, Marcello &
Bertoldi, Nicola & Cowan, Brooke & Shen, Wade & Moran, Christine & Zens, Richard & Dyer,

47
Chris & Bojar, Ondrej & Constantin, Alexandra & Herbst, Evan, (2007) “Moses: Open Source
Toolkit for Statistical Machine Translation”, ACL.
[22] Koehn, Philipp & Och, Franz.J & Marcu, Daniel, (2003) “Statistical Phrase-Based Translation”,
NAACL-HLT.
[23] Mikolov, Tomas & Karafiat, Martin & Burget, Lukas & Cernocky, Jan & Khudanpur, Sanjeev, (2010)
“Recurrent neural network based language model”, Interspeech.
[24] Papineni, Kishore & Roukos, Salim & Ward, Todd Zhu, Wei.J, (2002) “BLEU: a Method for
Automatic Evaluation of Machine Translation”.
[25] Sennrich, Rico & Haddow, Barry & Brich, Alexandra, (2016) “Improving Neural Machine
Translation Models with Monolingual Data”.
[26] Sutskever, Ilya & Vinyals, Oriol & Le, Quoc.V, (2014) “Sequence to Sequence Learning with Neural
Networks”.
[27] Tiedemann, Jorg, (2012) “Parallel Data, Tools and Interfaces in OPUS”, LREC.
[28] Utiyama, Masao & Isahara, Hitoshi, (2007) “A comparison of pivot methods for phrase-based
statistical machine translation”, NAACL-HLT.
[29] Wu, Hua & Wang, Haifeng, (2007) “Pivot language approach for phrase-based statistical machine
translation”, ACL.
[30] Zhang, Jiajun & Zong, Chengqing, (2016) “Exploiting source-side monolingual datain neural machine
translation”, EMNLP.
[31] Zoph, Barret & Yuret, Deniz & May, Jonathan & Knight, Kevin, (2016) “Transfer learning for low-
resource neural machine translation”, EMNLP.

EMPLOYING PIVOT LANGUAGE TECHNIQUE THROUGH STATISTICAL AND NEURAL MACHINE TRANSLATION FRAMEWORKS: THE CASE OF UNDER-RESOURCED PERSIAN-SPANISH LANGUAGE PAIR

More Related Content

What's hot (19)

Similar to EMPLOYING PIVOT LANGUAGE TECHNIQUE THROUGH STATISTICAL AND NEURAL MACHINE TRANSLATION FRAMEWORKS: THE CASE OF UNDER-RESOURCED PERSIAN-SPANISH LANGUAGE PAIR (20)

Recently uploaded (20)

EMPLOYING PIVOT LANGUAGE TECHNIQUE THROUGH STATISTICAL AND NEURAL MACHINE TRANSLATION FRAMEWORKS: THE CASE OF UNDER-RESOURCED PERSIAN-SPANISH LANGUAGE PAIR