Augmented Spanish-Persian Neural Machine Translation
Benyamin Ahmadnia and Raul Aranovich
Department of Linguistics, University of California at Davis, U.S.A.
Keywords:
Computational Linguistics, Natural Language Processing, Machine Translation, Low-resource Language
Pairs.
Abstract:
Neural Machine Translation (NMT) performs training of a neural network employing an encoder-decoder
architecture. However, the quality of the neural-based translations predominantly depends on the availability
of a large amount of bilingual training dataset. In this paper, we explore the performance of translations
predicted by attention-based NMT systems for Spanish to Persian low-resource language pairs. We analyze
the errors of NMT systems that occur in the Persian language and provide an in-depth comparison of the
performance of the system based on variations in sentence length and size of the training dataset. We evaluate
our translation results using BLEU and human evaluation measures based on the adequacy, fluency, and overall
rating.
1 INTRODUCTION
Complexity of the Statistical Machine Translation
(SMT) (Koehn et al., 2003) paradigm due to train-
ing of multiple components and inability to capture
long-term dependencies has diverted the attention of
researchers to neural-based approaches. With the evo-
lution of Neural Machine Translation (NMT) (Bah-
danau et al., 2015) as a modern methodology for
translation, MT systems have proved to exhibit fur-
ther improvements over the SMT systems. The en-
coder and decoder form the central units of the NMT
systems, providing an end-to-end system for transla-
tion from a source natural language to the target one
(Cho et al., 2014). The encoder performs transfor-
mation of the source sentences of variable length into
fixed length vectors, and the decoder then, generates a
variable-length output from the fixed representations.
Basically, NMT systems were developed based on
sequence to sequence learning on languages utiliz-
ing a gated Recurrent Neural Network (RNN) which
comprises of Long Short-Term Memory (LSTM) or
Gated Recurrent Unit (GRU) encoder-decoder ar-
chitecture. Eventually, improvements on the NMT
model was achieved using a LSTM with “global”
and “local” attention-based mechanisms for both the
encoder and decoder (Luong et al., 2015). Global
attention-based LSTM considers all the source posi-
tions at each timestamp whereas local attention at-
tends only to a subset of the source position. How-
ever, much of the positive outcome of NMT sys-
tems owe to the availability of large bilingual training
datasets. Consequently, in the case of low-resource
languages, there is a concern on the performance of
NMT systems being particularly lower as compared
to languages with largely available parallel corpus.
In this paper, we investigate the effectiveness of
NMT systems on Spanish-Persian low-resource trans-
lation. Our motivation for choosing Spanish and Per-
sian as the case-study is the linguistic differences
between these languages, which are from different
language families and have significant differences in
their properties, may pose a challenge for MT.
Low-resource languages, also known as resource
poor, are those that have fewer technologies and
datasets relative to some measure of their interna-
tional importance. In simple words, the languages
for which parallel training data is extremely sparse,
requiring recourse to techniques that are complemen-
tary to standard MT approaches. The biggest issue
with low-resource languages is the extreme difficulty
of obtaining sufficient resources.
Natural Language Processing (NLP) methods that
have been created for analysis of low-resource lan-
guages are likely to encounter similar issues to those
faced by documentary and descriptive linguists whose
primary endeavor is the study of minority languages.
Lessons learned from such studies are highly informa-
tive to NLP researchers who seek to overcome anal-
ogous challenges in the computational processing of
these types of languages (Ahmadnia and Dorr, 2019).
Availability of Spanish-Persian parallel data in
482
Ahmadnia, B. and Aranovich, R.
Augmented Spanish-Persian Neural Machine Translation.
DOI: 10.5220/0010369804820488
In Proceedings of the 13th International Conference on Agents and Artificial Intelligence (ICAART 2021) - Volume 1, pages 482-488
ISBN: 978-989-758-484-8
Copyright
c
2021 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
digital form is limited and inadequate for perform-
ing data-driven translations. Most of the data avail-
able online are not suitable for direct usage in prepa-
ration of the corpus. The data first needs to be cleaned
and carefully preprocessed for use in research which
is both time-consuming and tedious. To address this
issue, our goal is to prepare a high-quality bilingual
corpora by collecting data covering various domains
from various online and offline sources for Spanish-
to-Persian translation.
We evaluate the NMT system predicted transla-
tions employing BLEU automatic measure and hu-
man evaluation measures; 1) adequacy, 2) fluency,
and 3) overall rating. Additionally, we perform a
comprehensive analysis on the errors in the predicted
translation based on best, average and worst perfor-
mance of test sentences. Performance of translated re-
sults of NMT system have been evaluated from differ-
ent aspects, based on the variations in sentence length
and size of training data.
This paper is organized as follows; Section 2, re-
views our caste-study language issues. Section 3 de-
scribes the relevant researches related to MT, partic-
ularly in low-resource conditions. Section 4 details
the architecture of NMT system. Section 5 outlines
the experimental design and corpus description. Sec-
tion 6 describes analyzes the results. Section 7 con-
cludes the paper.
2 PERSIAN VS. SPANISH
Persian significantly suffers from shortage of digitally
available parallel and monolingual texts. It is mor-
phologically rich, with many characteristics shared
by Urdu and Arabic. It makes no use of articles (a,
an, the) and no distinction between capital and lower-
case letters. Symbols and abbreviations are rarely
used. As a consequence of being written in the Ara-
bic script, Persian uses a set of diacritic marks to indi-
cate vowels, which are generally omitted except in in-
fant writing or in texts for those who are learning the
language. Sentence structure is also different from
that of English. Persian places parts of speech such
as nouns, adverbs, and verbs in different locations in
the sentence, and sometimes even omits them alto-
gether. Some Persian words have many different ac-
cepted spellings, and it is not uncommon for trans-
lators to invent new words. This can result in OOV
words.
Spanish utilizes the Latin alphabet, with a few
special letters, vowels with an acute accent (
´
a,
´
u,
´
e,
´
o,
´
ı), u with an umlaut (
¨
u), and an n with a tilde (
˜
n). Due
to a number of reforms, the Spanish spelling system
is almost perfectly phonemic and, therefore, easier to
learn than the majority of languages. Spanish is pro-
nounced phonetically, but includes the trilled r which
is somewhat complex to reproduce. In the Spanish
IPA, the letters b and v correspond to the same symbol
b and the distinction only exists in regional dialects.
The letter h is silent except in conjunction with c, ch,
which changes the sound into tf. Spanish language
punctuation is very close to English. There are a few
significant differences; For example, in Spanish, ex-
clamative and interrogative sentences are preceded by
inverted question and exclamation marks. Also, in a
Spanish conversation, a change in speakers is indi-
cated by a dash, while in English, each speaker’s re-
mark is placed in separate paragraphs. Formal and
informal translations address several different charac-
teristics. Inflection, declination and grammatical gen-
der are important features of Spanish language.
A number of divergences (Dorr, 1994; Dorr et al.,
2002) between low-resource (e.g., Persian) and high-
resource (e.g., Spanish) languages pose many chal-
lenges in translation. In Persian, the modifier pre-
cedes the word it modifies, and in Spanish the modi-
fier follows the head word (although it may precede
the head word under certain conditions). In Per-
sian, sentences follow a “Subject”, “Object”, “Verb”
(SOV) order, and in Spanish, the sentences follow the
“Subject”, “Verb”, “Object” (SVO) order (Ahmadnia
et al., 2017). Such distinctions are exceedingly preva-
lent and thus pose many challenges for machine trans-
lation.
3 RELATED WORK
In the case of low-resource language settings, NMT
models have already been explored to perform com-
paratively poorer than statistical models owing to the
large parameter space of neural models and are often
prone to overfitting.
Employing transfer learning, performance of a
low-resource language pair has been improved by
transferring the learned parameters from a set of high-
resource languages to a set of low-resource languages.
A comparison of English-Basque translation has been
carried out using Google translate, NMT through
OpenNMT (Kelin et al., 2017) and SMT through
Moses (Koehn et al., 2007). Both NMT and SMT out-
performed Google Translate when instances of train-
ing and testing sets were used from the same cor-
pora. However, Google Translate performed better
on longer, more complex sequences (Unanue et al.,
2018).
One of the common approaches for dealing with
Augmented Spanish-Persian Neural Machine Translation
483
low-resource language pairs is employing monolin-
gual data to enhance the translation prediction. Ac-
cording to the supervised learning method, limited
parallel sentences are used along with the monolin-
gual data for translation (G
¨
ulc¸ehre et al., 2015; He
et al., 2016; Gu et al., 2018a; Ahmadnia and Dorr,
2019).
Using unsupervised learning (Lample et al.,
2018), variants of NMT and SMT models has been
proposed using three principles of initializing param-
eters, language models and back-translation to gen-
erate parallel text. However, the semi-supervised
learning method outperforms the other baseline ap-
proaches. According to Model-Agnostic Meta-
Learning (MAML) (Finn et al., 2017), low-resource
languages using the universal lexical representation
technique makes effective use of the availability
of high-resource languages pairs indeed (Gu et al.,
2018b).
NMT models are proposed by jointly incorporat-
ing RNN and Convolutional Neural Network (CNN)
in order to handle longer sequences more efficiently
based on the time-step and mini-batch dimensions.
In the absence of a parallel corpora or for zero-
resource MT, various approaches have already been
proposed which are broadly classified as multilingual
and pivot-based approaches (Ahmadnia et al., 2017).
Using a multilingual approach, focus is on exploit-
ing a multilingual parallel data for the translation. A
multi-way, multilingual NMT has been proposed for
zero-resource language where using a many-to-one
strategy was found to perform better than a one-to-
one strategy for zero-resource translation (Firat et al.,
2016).
An NMT system for translation of multilingual
languages has been proposed which uses the same
single NMT architecture model (Johnson et al., 2017).
It introduces the use of a token that indicates the target
language at the beginning of each input source that
enables it to perform multilingual translation. An-
other approach uses a pivot or a third language to help
with the translation in the absence of parallel corpora
(Chen et al., 2017). These approaches achieve zero re-
source translation effectively, however, difficulty lies
in universal representation for multiple languages as
well as increased complexity of the NMT models.
4 TRANSLATION SYSTEM
ARCHITECTURE
In this paper, we used a global attention-based
bidirectional LSTM on top of OpenNMT
1
as an
open-source library that implements the sequence-to-
sequence model with attention mechanism and per-
formed preprocessing, training, and translation on the
test dataset.
OpenNMT employs sequence-to-sequence mod-
els and supports attention mechanism. This open-
source library includes vanilla NMT models as well
as attention mechanism, gating, input feeding, regu-
larization, beam-search, etc. It also supports various
modules such as encoder, decoder, embedding layer,
embeddings, etc.
Our model employs an encoder with a stacked 2-
layer RNN with LSTM consisting of 500 hidden units
for training. The encoder reads all the source indi-
vidual word until it comes across the end-of-sentence
symbol (<eos>) and transforms the variable-length
input to a fixed vector representation known as “1-
hot” encoded vector. The length of such a vector
equals to the size of the vocabulary which represents
a vector of zeros where only the bit position of the
source word in the vocabulary is set to 1. Such a rep-
resentation takes up a lot of memory and is inefficient
if the size of the vocabulary is large. In order to de-
crease the size of the vector and to capture words with
similar context, embedding is employed.
Similar to encoder, the decoder is a 2-layer RNN
with LSTM consisting of 500 hidden units. At-
tending to all source hidden states of the input se-
quence, global attention is used at each time-step. An
alignment vector (a
t
) is computed by estimating each
source hidden vector (h
0
s
) with the target hidden vec-
tor (h
t
):
a
t
=
exp (score (h
t
, h
0
s
))
∑
s
0
exp (score (h
t
, h
0
s
0
))
(1)
The size of (a
t
) is equal to the number of source se-
quences. The score function is calculated as follows:
score(h
t
, h
0
s
) = h
t
W
a
h
0
s
(2)
The (a
t
) along with (h
0
s
) is used to derive context vec-
tor (C
t
) by taking a weighted average on all the source
side hidden states. A concatenation of (C
t
) and (h
t
)
then gives the attentional hidden vector (h
0
t
) as fol-
lows:
h
0
t
= tanh(W
c
[C
t
, h
t
]) (3)
A softmax layer is finally applied to vector (h
0
t
) in or-
der to produce the translated sentence in the target
language.
1
http://opennmt.net/
NLPinAI 2021 - Special Session on Natural Language Processing in Artificial Intelligence
484
Generally, translation is carried out using a beam
search to figure out the most adequate translation
from a set of all possible candidate translations. At
each level, the beam search provides a predetermined
number of most probable candidate results specified
by a beam width parameter which is set to a smaller
value for our experiment. A larger value of beam
width will be more likely to produce high-quality out-
put sequences but at the cost of translation time and
search accuracy. A beam width of size 1 would be
similar to a greedy approach that selects the most
likely target word. The beam search also supports
a number of normalization techniques such as length
normalization, coverage normalization, and end of
sentence normalization.
5 EXPERIMENTAL
FRAMEWORK
For the Spanish-Persian translation, we created a
bilingual corpus from various sources. The cor-
pus consists of parallel pairs of source and target
sentences that span multiple domains (GNOME
2
,
Tanzil
3
, OpenSubtitles2018
4
). Table 1 shows the data
statistics:
To prepare the training corpus, we performed
space-based tokenization on the text for separating
the tokens. Our dataset was separated into source and
target training and validation sets. The validation set
helps in the selection of models during training. The
preprocessing step builds dictionaries for mapping the
source and target vocabularies to their corresponding
indices. It then performs shuffling of the input data
so that each batch contains sentences from different
parts of the corpora. Sorting carried out so that sen-
tences of similar lengths are grouped together. Sen-
tences that are longer than a certain threshhold (set to
80) are discarded during the preprocessing stage.
A 2-layer (bidirectional) LSTM model has been
trained on our parallel corpora. The trained models
obtained for a total of 100 epochs
5
. The model gen-
erated at each epoch is used to translate our test data
and BLEU (Papineni et al., 2001) score is computed
for each translated output. This iterative process al-
lows us to find the most optimal epoch model (highest
BLEU score) among all 100 epoch models and to an-
2
http://opus.nlpl.eu/GNOME-v1.php
3
http://opus.nlpl.eu/Tanzil-v1.php
4
http://opus.nlpl.eu/OpenSubtitles-v2018.php
5
An epoch is when the entire corpus passes through the
neural network exactly once.
alyze each model for its effectiveness in translation
6
.
As an optimization method, we used the Stochastic
Gradient Descent (SGD) and learning rate set to 1.
The initial learning rate has been decayed by a factor
of 0.5 if no decrease in the validation perplexity oc-
curred after the ninth epoch. The learning rate was
decayed at each epoch thereafter.
To assess the quality of the predicted translation,
we used both automatic and human evaluation. For
automatic evaluation, BLEU metric (up to 3-grams
precision) was employed. On the other hand, human
evaluation was performed by two experts in the target
language, based on the adequacy, fluency, and overall
rating of each predicted output.
To further analyze of the translation results on the
basis of the length of the sentences, we grouped sen-
tences from the test set into three disjoint sets forming
three sentence groups: 1) Test1 (sentences of word
length 1-5), 2) Test2 (sentences of word length 6-10),
and 3) Test3 (sentence of word length greater than
10). Each of these sentence groups consists of 80 sen-
tences. For each sentence group, the BLEU scores
computed to evaluate the performance of the NMT
system based on variation in length of the source sen-
tences.
6 RESULTS ANALYSIS
We trained the system for a total of 100 epochs and
determined the BLEU scores (1-gram precision) for
each epoch. Since SGD is used as an optimiza-
tion method, the iterative method tends to produce an
under-fitted model with smaller number of training
epochs. However, the learning rate is decayed after
the ninth epoch with a decay factor of 0.5 after every
epoch to obtain the optimal model. We obtained the
highest BLEU score (1-gram precision) of 41.53 at
the 25th epoch which was considered as the optimal
epoch model for our system.
We computed the BLEU scores obtained at the
25th epoch model (the best performing model) for
1-gram, 2-grams, and 3-grams precisions and av-
erage BLEU scores for the test sets (see Table 2).
The NMT system obtained the highest average-BLEU
score for Test2 dataset and lowest average-BLEU
score for Test1 dataset. The higher BLEU score of
Test3 dataset compared to Test1 can be attributed to
the fact that although the former consists of complex,
compound and long sentences, the sentences being
held out from the train data have a similar structure
to the sentences in the training corpus.
6
The NMT system is trained on a single GPU
Augmented Spanish-Persian Neural Machine Translation
485
Table 1: Data statistics for Spanish-Persian translation.
Corpus Training Validation Test1 Test2 Test3
Tanzil 50K 1K 1.5K 1.5K 1.5K
GNOME 100K 1.5K 2K 2K 2K
OpenSubtitles2018 150K 2K 2.5K 2.5K 2.5K
Total 300K 4.5K 6K 6K 6K
Table 2: Spanish-Persian translation results based on
BLEU scores.
Test dataset 1-gram 2-grams 3-grams
Test1 41.53 27.54 20.79
Test2 44.39 34.08 28.81
Test3 43.77 33.25 25.68
Table 3: Spanish-Persian human evaluation; Fluency.
Test dataset Expert1 Expert2 Average
Test1 4.32 5.54 4.93
Test2 4.37 5.59 4.98
Test3 4.24 4.72 4.48
Furthermore, we observed that the NMT system per-
forms better on short sentences as compared to long
sentences. The observation is consistent with previ-
ous observations in (Bahdanau et al., 2015; Zhang
et al., 2017). This can attributed to the inability of
LSTM to capture long term dependencies very well,
particularly in low-resource conditions. Additionally,
we computed the average length of the translated or
predicted output for each sentence group. The length
of the translated output remains more or less within
the same group length but tend to produce shorter
translations on the average.
Two human experts fluent in the target language
evaluated the various test sets by rating each target or
predicted translation against the reference translation
on a predetermined scale of 1-5, where 5 indicates
highest score and 1 being the least. The ratings were
given based on three metrics;
• Adequacy: that shows how accurately the mean-
ing conveyed in a reference translation is pre-
served in the translated output.
• Fluency: that indicates the well-formedness of
the target translation.
• Overall Rating: that is given based on both of
the two metrics (adequacy and fluency) where a
high overall rating is assigned to target sentences
having high adequacy and fluency scores.
The average of the fluency, adequacy and overall rat-
ing scores of the two experts for each test set com-
puted and considered as the final scores. The experts
evaluated target translations predicted by the 25th
Table 4: Spanish-Persian human evaluation; Adequacy.
Test dataset Expert1 Expert2 Average
Test1 3.34 3.98 3.66
Test2 3.95 4.28 4.11
Test3 2.69 3.44 3.06
Table 5: Spanish-Persian human evaluation; Overall.
Test dataset Expert1 Expert2 Average
Test1 3.54 4.09 3.81
Test2 3.78 4.21 3.99
Test3 3.21 3.59 3.4
epoch model. Tables 3, 4, and 5 show the scores of
Fluency, Adequacy and Overall Rating respectively,
on the test sets.
All test sets achieve high fluency scores as op-
posed to the adequacy scores as NMT systems tend
to produce syntactically correct translations. The av-
erage score ratings for BLEU and human evaluation
has a negative correlation, with high BLEU score and
low human evaluation score. The higher BLEU score
can be attributed to the principle of BLEU metric.
BLEU uses n-gram precision and for long sentences
in Test3, intuitively, the n-gram matches are higher
for long sentences as compared to short sentences in
the other two test sets. However, although some lexi-
cal similarities may exist between reference and pre-
dicted translations, BLEU does not take into account
the syntactic and semantic constructs of the predicted
translation.
Figures 1 to 4 show some examples from the
candidate translations predicted by best performing
epoch model and classified them into best, average
and worst performance based on the overall rating of
the human evaluation results to analyze the quality of
the results of translations predicted by the NMT sys-
tem.
In Figure 1, the NMT system correctly translates
the source sentence and the predicted output achieves
an overall rating of 5 which means it is perfectly ade-
quate and fluent. In Figure 2, NMT system translates
the source sentence to “thank you for your advice”.
There is a mistranslation of the word “ayuda” which
is partially adequate but completely fluent translation.
Although the predicted output is completely inade-
NLPinAI 2021 - Special Session on Natural Language Processing in Artificial Intelligence
486
Source:
¿Cmo sueles pasar tu tiempo en casa?
Reference:
Predicted output:
Source:
Que Dios te bendiga a ti y a tu familia
Reference:
Predicted output:
Source:
Gracias por su ayuda
Reference:
Predicted output:
Source:
Prometo que no lo volvería a hacer
Reference:
Predicted output:
Figure 1: The best translation performance.
Source:
¿Cmo sueles pasar tu tiempo en casa?
Reference:
Predicted output:
Source:
Que Dios te bendiga a ti y a tu familia
Reference:
Predicted output:
Source:
Gracias por su ayuda
Reference:
Predicted output:
Source:
Prometo que no lo volvería a hacer
Reference:
Predicted output:
Figure 2: The worst translation performance.
Source:
¿Cmo sueles pasar tu tiempo en casa?
Reference:
Predicted output:
Source:
Que Dios te bendiga a ti y a tu familia
Reference:
Predicted output:
Source:
Gracias por su ayuda
Reference:
Predicted output:
Source:
Prometo que no lo volvería a hacer
Reference:
Predicted output:
Figure 3: The best translation performance.
Source:
¿Cmo sueles pasar tu tiempo en casa?
Reference:
Predicted output:
Source:
Que Dios te bendiga a ti y a tu familia
Reference:
Predicted output:
Source:
Gracias por su ayuda
Reference:
Predicted output:
Source:
Prometo que no lo volvería a hacer
Reference:
Predicted output:
Figure 4: Adequate translation performance.
quate, the NMT system still produces a fluent transla-
tion. In Figure 3, predicted output by the NMT system
is perfectly fluent and adequate. The overall rating
given by all the three evaluators is 5. In Figure 4, the
predicted output is perfectly fluent and almost ade-
quate as the phrase “en casa” is omitted in the transla-
tion predicted by the NMT system achieving an over-
all rating of 4.
According to the above example of best, average,
and worst performance translations, our observations
are listed as follows:
• In most cases, NMT systems tends to generate
fluent and syntactically correct translations, even
in instances where the predicted output is com-
pletely inadequate.
• However, the fluency of NMT systems reduces
with increase in the length of the sentences and
with complex and compound sentences which
consists of independent or dependent clauses.
This may be attributed to the inability of NMT
systems to capture long-term dependencies.
• NMT systems often mistranslate named entities
such as person names, location as well as num-
bers which reduces the preciseness and accuracy
of the translation. Often, NMT systems trades off
adequacy for fluency of the predicted translations.
• Although NMT systems use input feeding mecha-
nism to keep track of past alignments, we observe
some source words have been translated multi-
ple times while other words have not been trans-
lated. This indicates that the use of input feeding
for alignment decision is not sufficient to prevent
over-translation or under-translation.
7 CONCLUSIONS
In this paper, we employed a global attentional NMT
system to train and test data for Spanish-to-Persian
language pair. The predicted translations evaluated
based on BLEU scores. Furthermore, human evalua-
tions based on adequacy and fluency were evaluated.
We also analyzed the performance of predicted trans-
lation based on various experimental settings. Our
analysis shows that the performance of NMT systems
increases with the increase of the size of training cor-
pus. We also observed that NMT systems often mis-
translate named entities and compromise adequacy
for fluency.
Although MT is unlikely to completely replace
human translations, the effectiveness of these auto-
matic translations have definitely surpassed profes-
sional human translators where quality can be com-
promised. Since the performance of NMT systems
are largely influenced by the size of the training cor-
pus, increasing the training corpus will undoubtedly
improve the translation results. Incorporating linguis-
tic resources such as monolingual data can also im-
prove translation results.
ACKNOWLEDGEMENTS
We thank the reviewers for valuable feedback and dis-
cussions. This work was supported by the Department
of Linguistics at UC Davis (USA).
REFERENCES
Ahmadnia, B. and Dorr, B. J. (2019). Augmenting neu-
ral machine translation through round-trip training ap-
proach. Open Computer Science, 9(1):268–278.
Ahmadnia, B., Serrano, J., and Haffari, G. (2017). Persian-
Spanish low-resource statistical machine translation
through english as pivot language. In Proceedings
of Recent Advances in Natural Language Processing,
pages 24–30.
Bahdanau, D., Cho, K., and Bengio, Y. (2015). Neural ma-
chine translation by jointly learning to align and trans-
late. In Proceedings of the International Conference
on Learning Representations.
Augmented Spanish-Persian Neural Machine Translation
487
Chen, Y., Liu, Y., Cheng, Y., and V. Li, V. (2017). A teacher-
student framework for zero-resource neural machine
translation. In Proceedings of the 55th Annual Meet-
ing of the Association for Computational Linguistics,
pages 1925–1935.
Cho, K., merrienboer, B. V., G
¨
ulc¸ehre, C¸ ., Bahdanau, D.,
Bougares, F., Schwenk, H., and Bengio, Y. (2014).
Learning phrase representations using rnn encoder-
decoder for statistical machine translation. In Pro-
ceedings of the conference on Empirical Methods in
Natural Language Processing, pages 1724–1734.
Dorr, B. J. (1994). Machine translation divergences: A
formal description and proposed solution. Computa-
tional Linguistics, 20(4):597–633.
Dorr, B. J., Pearl, L., Hwa, R., and Habash, N. (2002).
Duster: A method for unraveling cross-language di-
vergences for statistical word-level alignment. In Pro-
ceedings of the 5th conference of the Association for
Machine Translation in the Americas.
Finn, C., Abbeel, P., and Levine, S. (2017). Model-agnostic
meta-learning for fast adaptation of deep networks. In
Proceedings of the 34th International Conference on
Machine Learning, pages 1126–1135.
Firat, O., Sankaran, B., Al-onaizan, Y., Yarman Vural, F. T.,
and Cho, K. (2016). Zero-resource translation with
multi-lingual neural machine translation. In Proceed-
ings of the 2016 Conference on Empirical Methods in
Natural Language Processing, pages 268–277.
Gu, J., Hassan, H., Devlin, J., and Li, V. (2018a). Uni-
versal neural machine translation for extremely low
resource languages. In Proceedings of the 2018 Con-
ference of the North American Chapter of the Associa-
tion for Computational Linguistics: Human Language
Technologies, pages 344–354.
Gu, J., Wang, Y., Chen, Y., Li, V., and Cho, K. (2018b).
Meta-learning for low-resource neural machine trans-
lation. In Proceedings of the 2018 Conference on
Empirical Methods in Natural Language Processing,
pages 3622–3631.
G
¨
ulc¸ehre, C¸ ., Firat, O., Xu, K., Cho, K., Barrault, L.,
Lin, H., Bougares, F., Schwenk, H., and Bengio, Y.
(2015). On using monolingual corpora in neural ma-
chine translation. ArXiv, abs/1503.03535.
He, D., Xia, Y., Qin, T., Wang, L., Yu, N., Liu, T., and Ma,
W. (2016). Dual learning for machine translation. In
Proceedings of the 30th Conference on Neural Infor-
mation Processing Systems.
Johnson, M., Schuster, M., Le, Q. V., Krikun, M., Wu, Y.,
Chen, Z., Thorat, N., Vi
´
egas, F., Wattenberg, M., Cor-
rado, G., Hughes, M., and Dean, J. (2017). Google’s
multilingual neural machine translation system: En-
abling zero-shot translation. Transactions of the Asso-
ciation for Computational Linguistics, 5:339–351.
Kelin, G., Kim, Y., Deng, Y., Senellart, J., and Rush, A. M.
(2017). OpenNMT: Open-source toolkit for neural
machine translation. In Proceedings of 55th Annual
Meeting of the Association for Computational Lin-
guistics, pages 67–72.
Koehn, P., Hoang, H., Birch, A., Callison-Burch, C., Fed-
erico, M., Bertoldi, N., Cowan, B., Shen, W., Moran,
C., and Zens, R. (2007). Moses: Open source toolkit
for statistical machine translation. In Proceedings of
the 45th annual meeting of the Association for Com-
putational Linguistics, pages 177–180.
Koehn, P., Och, F. J., and Marcu, D. (2003). Statistical
phrase-based translation. In Proceedings of the Con-
ference of the North American Chapter of the Associ-
ation for Computational Linguistics on Human Lan-
guage Technology, pages 48–54.
Lample, G., Ott, M., Conneau, A., Denoyer, L., and Ran-
zato, M. (2018). Phrase-based & neural unsupervised
machine translation. In Proceedings of the 2018 Con-
ference on Empirical Methods in Natural Language
Processing, pages 5039–5049.
Luong, T., Sutskever, I., Le, Q., Vinyals, O., and Zaremba,
W. (2015). Addressing the rare word problem in neu-
ral 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, pages 11–19.
Papineni, K., Roukos, S., Ward, T., and Zhu, W. (2001).
Bleu: A method for automatic evaluation of ma-
chine translation. In Proceedings of the 40th Annual
Meeting on Association for Computational Linguis-
tics, pages 311–318.
Unanue, I. J., Arratibel, L. G., Borzeshi, E. Z., and Piccardi,
M. (2018). English-basque statistical and neural ma-
chine translation. In Proceedings of the Eleventh In-
ternational Conference on Language Resources and
Evaluation.
Zhang, B., Xiong, D., Su, J., and Duan, H. (2017). A
context-aware recurrent encoder for neural machine
translation. IEEE/ACM Trans. Audio, Speech and
Lang. Proc., 25(12):2424–2432.
NLPinAI 2021 - Special Session on Natural Language Processing in Artificial Intelligence
488
