EXPERIMENTS ON DIFFERENT RECURRENT NEURAL NETWORKS FOR ENGLISH-HINDI MACHINE TRANSLATION

Dhinaharan Nagamalai et al. (Eds) : AIS, CSIT, IPPR, IPDCA - 2017
pp. 63– 74, 2017. © CS & IT-CSCP 2017 DOI : 10.5121/csit.2017.71006
EXPERIMENTS ON DIFFERENT
RECURRENT NEURAL NETWORKS FOR
ENGLISH-HINDI MACHINE
TRANSLATION
Ruchit Agrawal1
and Dipti Misra Sharma2
1
Language Technologies Research Center, IIIT Hyderabad
2
Head, Language Technologies Research Center, IIIT Hyderabad
ABSTRACT
Recurrent Neural Networks are a type of Artificial Neural Networks which are adept at dealing
with problems which have a temporal aspect to them. These networks exhibit dynamic
properties due to their recurrent connections. Most of the advances in deep learning employ
some form of Recurrent Neural Networks for their model architecture. RNN's have proven to be
an effective technique in applications like computer vision and natural language processing. In
this paper, we demonstrate the effectiveness of RNNs for the task of English to Hindi Machine
Translation. We perform experiments using different neural network architectures - employing
Gated Recurrent Units, Long Short Term Memory Units and Attention Mechanism and report
the results for each architecture. Our results show a substantial increase in translation quality
over Rule-Based and Statistical Machine Translation approaches.
KEYWORDS
Machine Translation, Recurrent Neural Networks, LSTMs, GRUs, English-Hindi MT.
1. INTRODUCTION
Deep learning is a rapidly advancing approach to machine learning and has shown promising
performance when applied to a variety of tasks like image recognition, speech processing, natural
language processing, cognitive modelling and so on. Deep Learning involves using large neural
networks for training a model for a specific task. This paper demonstrates the application of deep
learning for Machine Translation of English ! Hindi, two linguistically distant and widely spoken
languages. The application of deep neural networks to Machine Translation has been
demonstrated by (Kalchbrenner and Blunsom, 2013; Sutskever et al., 2014; Cho et al., 2014;
Bahdanau et al., 2014) and it has shown promising results for various language pairs.
In this paper, we experiment with different deep learning architectures. These include Gated
Recurrent Units (GRUs), Long Short Term Memory Units (LSTMs) and addition of attention
mechanism to each of these architectures. We demonstrate that the best performance for English -
> Hindi MT is generally obtained using Bi-directional LSTMs with attention mechanism and in

64 Computer Science & Information Technology (CS & IT)
some cases with GRUs with attention mechanism. The Bi-directional LSTMs generally show
better performance for compound sentences and larger context windows.
We show manual samples of output translations and provide their evaluation to demonstrate
the effectiveness of different architectures.
We describe the motivation behind the choice of RNNs in detail in Section 3. We briefly
discuss related work in Section 2, followed by the description of our neural network model in
Section 4. The experiments and results are discussed in Section 5. The paper is concluded in
Section 6.
2. RELATED WORK
The usage of large neural networks for Natural Language Processing (NLP) tasks was initially
proposed by (LeCun et al., 2015) in his feed-forward neural language model. The neural
Language Model he proposed is very similar to the current existing Language Models.
The input n-gram is projected into an embedding space for each word and passes to big output
layer.
Figure 1: A comparison of feedforward neural networks with Recurrent Neural Networks
This novel idea was then used by several researchers who tried to integrate it with Machine
Translation systems ((Auli et al., 2013) and (Cho et al., 2014)).
(Sutskever et al., 2014) was a breakthrough for Machine Translation, introducing the ”seq2seq”
(Sequence to sequence) model which was the first model based completely on neural networks
and achieving accuracy comparable to the State-of-the-Art SMT systems. They proposed the
usage of a Recurrent Neural Network model with the encoders and decoders comprising of
LSTMs or GRUs. They propose running the encoder over the source sentence, producing a
hidden state and then running another RNN (decoder) to generate the output one word at a time.
The bottleneck to this approach was that the entire translation is a fixed sized vector. There have
been different techniques (like padding) to rectify this issue.

Computer Science & Information Technology (CS & IT) 65
Anusaaraka (Bharati et al., 1994) is an English to Hindi Machine Translation, primarily Rule-
based, but employing a parser which uses statistical approaches (De Marneffe et al., 2006).
Figure 2: Structure of an LSTM unit
3. MOTIVATION BEHIND USING RECURRENT NEURAL NETWORKS
Traditional Neural Networks have a huge RAM requirement and are not quite feasible in their
best settings where they achieve their highest accuracies. Additionally, they are not designed to
deal with sequential information. We explain this below :
One important property of machine translation, or any task based on natural languages, is that we
deal with variable-length input and output. For example; if the input X=(x1; x2; : : : ; xT ) and
output Y=(y1; y2; : : : ; yT ′ ); The lengths of the sequences i.e. T and T’ are not fixed.
On the other hand, one of the major assumptions in feedforward neural networks is the idea of
fixed length, i.e. the size of the input layer is fixed to the length of the input sequence. The other
major assumption is the idea of independence - that different training examples (like images) are
independent of each other. However, we know of temporal sequences such as sentences or
speech, there are short and long temporal dependencies that have to be accounted for.
To deal with these types of variable-length input and output, we need to use a recurrent neural
network (RNN). Widely used feed-forward neural networks, such as convolutional neural
networks, do not maintain internal state other than the network’s own parameters. Whenever a
single sample is fed into a feed-forward neural network, the network’s internal state, or the
activations of the hidden units, is computed from scratch and is not influenced by the state
computed from the previous sample. On the other hand, an RNN maintains its internal state while
reading a sequence of inputs, which in our case will be a sequence of words, thereby being able to
process an input of any length.

Recurrent Neural Networks (RNNs) also address the independence issue - they facilitate the
preservation as well as processing of information that has a temporal aspect involved. For
example; a sequence of words has an order, and hence a time element inherent in it. A model
which takes this into consideration is needed for efficient performance. This is not possible if we
employ feed-forward neural networks. Thus, Recurrent Neural Networks can not only learn the
local and long term temporal dependencies in the data, but can also accommodate input
sequences of variable length.
The RNN’s thus help in converting the input sequence to a fixed size feature vector that encodes
primarily the information which is crucial for translation from the input sentence, and ignores the
irrelevant information. Figure 1 shows a comparison of feed-forward neural networks with
recurrent neural networks.
Long Short Term Memory (LSTM) units are a type of RNNs which are very good at preserving
information through time-steps over a period of time. Figure 2 shows the structure of an LSTM
unit. One key advance in LSTMs in recent years has been the concept of bi-directional encoder
and decoder framework. When we employ bidirectional LSTMs, we end up with two hidden
states - one in the forward direction and one in the backward direction. This allows the network to
learn from the text. Often, even more than two layers are used. Thus there will be multiple layers
stacked on top of each other - this is generally only in huge training data conditions. Each one of
these has a set of weights inside it, and learns and affects the one above it. The final state
represents everything that is in the source words. Bi-directional LSTMs generally work the best
specially when complemented with the attention mechanism.
After the encoding process, we are left with a context vector - which is like a snapshot of the
entire source sequence and is used further to predict the output. We have a dense layer with
softmax similar to a feed-forward neural network, but the difference is that it is time distributed
i.e. we have one of these for each time step. The top layer thus has one neuron for every single
word in the vocabulary and hence is huge in size in large vocabulary conditions.
Figure 3: A two-layered LSTM architecture which we employ in our experiments

4. FORMULATION OF OUR MODEL
In order to train the recurrent neural networks, we take the cost function and obtain its derivative
with respect to the weight in question. We then move this derivative through the nested layer of
computations using the chain rule.
In other words, the output of the previous layer is multiplied by the weight matrix and added to a
bias and then passed on to an activation function.
yk = g(W yk 1 + b) (1)
Table 1: Different Hindi translations corresponding to the English sentence - “Shyam has given the book to
Manish.” (Due to word order)
Table 2: Anusaaraka scores on ILCI test data
We use a recurrent connection convert the linear unit of feed-forward neural network to a
recurrent unit so that now the activity of the unit ht not only depends on xt (the input) multiplied
by the weight matrix, but also on its activity at the previous timestep. The following equation
shows this phenomenon :
h
(t)
= gh(W1x
(t)
+ WRh
(t 1)
+ bh) (2)
Table 3: Two different translations corresponding to the same English sentence - from ILCI test data
(Many-to-many mapping between vocabulary)

Table 4: Results - Comparison of metric scores obtained on two-layered and four-layered model at different
stages
The second term WRh
(t 1)
depends on the activity at the previous timestep multiplied by a
recurrent weight matrix. We also want to be able to retrieve an output from this unit and this is
done by adding a linear operation as described in the following equation :
y
(t)
= gy(Wyh
(t)
+ by) (3)
Here, y
(t)
is a function of h
(t)
multiplied by weight matrix w and passed through a non-linear
activation function. This is the basic element of the recurrent neuron which we use in our RNN
architectures.
The process can be visualized as the input sequence being compressed by the RNN into an
intermediate representation in the form of a fixed dimensional vector. So, if the vector ht 1
describes the history of the sequence at timestep t, the new internal state (the updated vector) ht
will be computed by the network, effectively compressing the preceding symbols (x1; x2; : : : ; xt
1) as well as the new symbol xt. The following equation shows this :
ht = ϕ (xt, ht 1)
Here, ϕ is a function which takes the new information unit xt and the hidden state ht 1 as input.
(h0 can be assumed to be a vector containing zeroes).

Table 5: Evaluating output quality : Different RNN architectures
Table 6: Evaluating output quality : Adding Attention Mechanism
5. EXPERIMENTS AND RESULTS
We employ a sequence-to-sequence model with Recurrent Neural Networks to train our models.
We conduct experiments on two and four layers of encoder and decoder respectively. We use the
architecture as described in Section 4. We use the seq2seq model available in Tensorflow1
to
implement the above mentioned architecture.
For training the model, we extract 200,000 sentences from the HindEnCorp (Bojar et al., 2014)
corpus. We employed pruning using appropriate rules to remove unsuitable sentences. For
-------
1
It can be accessed at : https://github.com/google/seq2seq

example, all sentences of length greater than fifty were removed from the corpus. The reason was
low scalability of neural networks to translate sentences of length greater than 50. Also, sentences
of length less than three were removed to discourage memorization, instead of syntactic and
semantic learning of concepts. Pruning was also done to remove special characters and
hyperlinks from the sentences.
After removing discrepancies, rest of the sentences were randomly shuffled to create the parallel
training corpus. We test the performance of our model using the ILCI test set (Jha, 2010) and the
WMT 2014 English-Hindi test set.
We observe that our model is able to produce grammatically fluent translations, as opposed to
traditional approaches. Some problems which still need to be solved are presence of repeated
tokens and unknown or unmapped words. A bi-directional LSTM model with attention
mechanism shows improvement over normal RNN’s in both these aspects.
Table 7: Evaluating output quality : Two layers vs. Four layers
Table 8: Results on WMT Test data

Table 4 demonstrates the performance of our model during various stages as measured by the
above-mentioned metrics. We observe on manual inspection of samples that there is a significant
improvement in performance over rule-based and statistical approaches by using deep neural
networks, thereby producing quality translation as shown by the use of semantically correct
synonyms. For example, Table 3 shows a sample sentence from the ILCI test corpus ( ILCItest )
and its corresponding output obtained by our model. The English as well as Hindi meaning of
both the sentences is the same, although they differ in their structure and words used in the Hindi
output. The LSTM output displays an impressive usage of the phrase “cAra cAzxa lagAwe hEM”
- a contextually suitable and semantically correct idiom in Hindi which conveys “enhancing of
personality”.
Anusaaraka has a BLEU score of 6:98 on ILCI test data (Table 2). We observe a 4:72 point
increase in the BLEU score by using GRUs. Similar improvements can be seen for other metrics
by using different RNN architectures. Table 5 shows the variation in quality of translation
obtained on using different RNN architectures. The Anusaaraka output does not make much sense
(is syntactically as well as semantically poor) and the GRU a grammatically incorrect sentence.
While the LSTM model produces a better translation with a minor error in pronoun usage, the Bi-
directional LSTM model generates the correct output.
We demonstrate the effect of addition of attention mechanism in Table 6. Table 7 compares the
output of two-layered model and four-layered model obtained on the different architectures using
sample translations. We can observe that the four-layered model is able to perform better in
many cases two-layered counterpart. The reason can be attributed to higher complexity of this
model and sufficient data for training.
We also conduct experiments and report results on the WMT-14 corpus in Table 8. The results
further improve on using Bi-directional LSTM with attention to give a BLEU score of 9.23,
comparable to (Dungarwal et al., 2014), a system fully trained on the WMT training corpus.
6. CONCLUSION AND FUTURE WORK
In this paper, we build sequence-to-sequence models using Recurrent Neural Networks. We
experimented with Gated Recurrent Units, Long Short Term Memory Units and the attention
mechanism. We demonstrated results using this approach on a linguistically distant language pair
En ! Hi and showed a substantial improvement in translation quality. We conclude that Recurrent
Neural Networks perform well for the task of English-Hindi Machine Translation. The bi-
directional LSTM units perform best, specially on compound sentences. Future work includes
performing experiments on other languages, specially among morphologically rich languages,
like Indian to Indian language MT. We would like to explore MT for resource-scarce languages,
in conditions where large parallel corpora for training are not available.
REFERENCES
[1] Gary Anthes. 2010. Automated translation of indian languages. Communications of the ACM
53(1):24– 26.
[2] Michael Auli, Michel Galley, Chris Quirk, and Geoffrey Zweig. 2013. Joint language and translation
modeling with recurrent neural networks. In EMNLP . volume 3, page 0.

[3] Akshar Bharati, Vineet Chaitanya, and Rajeev Sangal. 1994. Anusaraka or language accessor: A short
introduction. Automatic Translation, Thiruvananthpuram, Int. school of Dravidian Linguistics .
[4] Akshar Bharati, Vineet Chaitanya, Rajeev Sangal, and KV Ramakrishnamacharyulu. 1995. Natural
language processing: a Paninian perspective . Prentice-Hall of India New Delhi.
[5] Roger M Blench and M Post. Rethinking sino-tibetan phylogeny from the perspective of north east
indian languages. paper accepted for a volume of selected papers from the 16th himalayan languages
symposium 2-5 september 2010 school of oriental and african studies, london. ed. Nathan Hill.
Mouton de Gruyter .
[6] Ondrej Bojar, Vojtech Diatka, Pavel Rychlỳ, Pavel Stranák, Vít Suchomel, Ales Tamchyna, and
Daniel Zeman. 2014. Hindencorp-hindi-english and hindi-only corpus for machine translation. In
LREC . pages 3550–3555.
[7] Kyunghyun Cho, Bart Van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares,
Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using rnn encoder-
decoder for statistical machine translation. arXiv preprint arXiv:1406.1078 .
[8] Junyoung Chung, Caglar Gülçehre, Kyunghyun Cho, and Yoshua Bengio. 2015. Gated feedback
recurrent neural networks. In ICML . pages 2067–2075.
[9] Marie-Catherine De Marneffe, Bill MacCartney, Christopher D Manning, et al. 2006. Generating
typed dependency parses from phrase structure parses. In Proceedings of LREC . Genoa Italy, volume
6, pages 449–454.
[10] George Doddington. 2002. Automatic evaluation of machine translation quality using n-gram co-
occurrence statistics. In Proceedings of the second international conference on Human Language
Technology Research . Morgan Kaufmann Publishers Inc., pages 138–145.
[11] Piyush Dungarwal, Rajen Chatterjee, Abhijit Mishra, Anoop Kunchukuttan, Ritesh Shah, and
Pushpak Bhattacharyya. 2014. The iit bombay hindi english translation system at wmt 2014. ACL
2014 page 90.
[12] Aria Haghighi and Dan Klein. 2009. Simple coreference resolution with rich syntactic and semantic
features. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language
Processing: Volume 3-Volume 3 . Association for Computational Linguistics, pages 1152–1161.
[13] Hideki Isozaki, Tsutomu Hirao, Kevin Duh, Katsuhito Sudoh, and Hajime Tsukada. 2010. Automatic
evaluation of translation quality for distant language pairs. In Proceedings of the 2010 Conference on
Empirical Methods in Natural Language Processing . Association for Computational Linguistics,
pages 944–952.
[14] Girish Nath Jha. 2010. The tdil program and the indian langauge corpora intitiative (ilci). In LREC .
[15] Nal Kalchbrenner and Phil Blunsom. 2013. Recurrent continuous translation models. In EMNLP . 39,
page 413.
[16] Nayan Jyoti Kalita and Baharul Islam. 2015. Bengali to assamese statistical machine translation using
moses (corpus based). arXiv preprint arXiv:1504.01182 .

[17] Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola
Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, et al. 2007. Moses: Open
source toolkit for statistical machine translation. In Proceedings of the 45th annual meeting of the
ACL on interactive poster and demonstration sessions . Association for Computational Linguistics,
pages 177–180.
[18] Alon Lavie and Michael J Denkowski. 2009. The meteor metric for automatic evaluation of machine
translation. Machine translation 23(2):105–115.
[19] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. 2015. Deep learning. Nature 521(7553):436–
444.
[20] Minh-Thang Luong and Christopher D Manning. 2015. Stanford neural machine translation systems
for spoken language domains. In Proceedings of the International Workshop on Spoken Language
Translation .
[21] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. 2015. Effective approaches to
attention-based neural machine translation. arXiv preprint arXiv:1508.04025 .
[22] Anthony McEnery, Paul Baker, Rob Gaizauskas, and Hamish Cunningham. 2000. Emille: Building a
corpus of south asian languages. VIVEK-BOMBAY- 13(3):22–28.
[23] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic
evaluation of machine translation. In Proceedings of the 40th annual meeting on association for
computational linguistics . Association for Computational Linguistics, pages 311–318.
[24] Reinhard Rapp and Carlos Martin Vide. 2006. Example-based machine translation using a dictionary
of word pairs. In Proceedings, LREC . pages 1268–1273.
[25] Alexander J Ratner, Christopher M De Sa, Sen Wu, Daniel Selsam, and Christopher Ré. 2016. Data
programming creating large training sets, quickly. In Advances in Neural Information Processing
Systems . pages 3567–3575.
[26] Matthew Snover, Bonnie Dorr, Richard Schwartz, Linnea Micciulla, and John Makhoul. 2006. A
study of translation edit rate with targeted human annotation. In Proceedings of association for
machine translation in the Americas . volume 200.
[27] Martin Sundermeyer, Ralf Schlüter, and Hermann Ney. 2012. Lstm neural networks for language
mod-eling. In Interspeech . pages 194–197.
[28] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural
networks.
[29] In Advances in neural information processing systems . pages 3104–3112.
[30] Nicola Ueffing, Gholamreza Haffari, Anoop Sarkar, et al. 2007. Transductive learning for statistical
machine translation. In Annual Meeting-Association for Computational Linguistics . volume 45, page
25.
[31] Paul J Werbos. 1990a. Backpropagation through time: what it does and how to do it. Proceedings of
the IEEE 78:1550–1560.
[32] Paul J Werbos. 1990b. Backpropagation through time, what it does and how to do it. Proceedings of
the IEEE 78.

[33] David Yarowsky. 1995. Unsupervised word sense disambiguation rivaling supervised methods. In
Pro-ceedings of the 33rd annual meeting on Association for Computational Linguistics . Association
for Computational Linguistics, pages 189–196.

EXPERIMENTS ON DIFFERENT RECURRENT NEURAL NETWORKS FOR ENGLISH-HINDI MACHINE TRANSLATION

More Related Content

What's hot (20)

Similar to EXPERIMENTS ON DIFFERENT RECURRENT NEURAL NETWORKS FOR ENGLISH-HINDI MACHINE TRANSLATION (20)

Recently uploaded (20)

EXPERIMENTS ON DIFFERENT RECURRENT NEURAL NETWORKS FOR ENGLISH-HINDI MACHINE TRANSLATION