Available online at www.sciencedirect.

com
ScienceDirect
ScienceDirect
Procedia Computer Science 00 (2019) 000–000
Available online at www.sciencedirect.com
Procedia Computer Science 00 (2019) 000–000 www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia
ScienceDirect
Procedia Computer Science 167 (2020) 2534–2544

International Conference on Computational Intelligence and Data Science (ICCIDS 2019)

International Conference on Computational Intelligence and Data Science (ICCIDS 2019)
Corpus based Machine Translation System with Deep Neural
Corpus based Machine
Network Translation
for Sanskrit System
to Hindi with Deep Neural
Translation
Network for Sanskrit to Hindi Translation
Muskaan Singha,Ravinder Kumara,Inderveer Chanab*
Muskaan Singha,Ravinder Kumara,Inderveer Chanab*
a
Language Engineering and Machine Learning Research Labs, Thapar Institute of Engineering and Technology,Patiala,India.
a
Language Engineering andCSED, Thapar
Machine Institute
Learning of Engineering
Research and Technology,
Labs, Thapar Patiala, India.
Institute of Engineering and Technology,Patiala,India.
CSED, Thapar Institute of Engineering and Technology, Patiala, India.

Abstract
Abstract
Sanskrit language is the mother of almost all Indian languages. The main requirement in Sanskrit domain is to translate the life-
transforming
Sanskrit stories
language (epics),
is the Vedas
mother etc. toall
of almost make them
Indian availableThe
languages. in other
main languages,
requirementforin public
Sanskritat domain
large. Inisthe field of machine
to translate the life-
translation system
transforming there
stories is need
(epics), to develop
Vedas etc. to amake
machine
themtranslation
available system
in otherwhich translate
languages, forSanskrit
public atlanguage
large. Intothe
Hindi.
fieldSo,
of the main
machine
focus of thissystem
translation work is to propose
there a new
is need to corpus-based
develop a machinetranslation
translationsystem
systemforwhich
Sanskrit to Hindi
translate translation
Sanskrit where
language Bhagvad
to Hindi. So,Gita – the
the main
song of
focus of the
thislord
workisisused as an input
to propose a newdata. In this work,
corpus-based Deep neural
translation systemnetwork is used
for Sanskrit for training
to Hindi where
translation input
where data is Gita
Bhagvad passed to
– the
neuralofnetwork
song the lordafter dataasanalysis
is used an inputand processing
data. whichDeep
In this work, then neural
performs auto-tuning
network is usedthat
forhelp to make
training wherethis model
input databetter. Target
is passed to
text is prepared
neural network using this proposed
after data model
analysis and and achieves
processing whichbetter
thenBLEU Score
performs and Wordthat
auto-tuning Error Rate.
help to make this model better. Target
text is prepared using this proposed model and achieves better BLEU Score and Word Error Rate.
© 2019 The Authors. Published by Elsevier B.V.. This is an open access article under the CC BY-NC-ND license
© 2020 The Authors. Published by Elsevier B.V.
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
© 2019 The Authors. Published by Elsevier B.V.. This is an open access article under the CC BY-NC-ND license
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the International Conference on Computational Intelligence and
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the International Conference on Computational Intelligence and Data
Data Science
Peer-review
Science (ICCIDS
under
(ICCIDS 2019)
responsibility
2019). of the scientific committee of the International Conference on Computational Intelligence and
Data Science (ICCIDS 2019)
Keywords: Corpus, MTS, Sanskrit, Hindi, Deep Neural Network.
Keywords: Corpus, MTS, Sanskrit, Hindi, Deep Neural Network.

1. Introduction
1. Introduction
The translation of one language to the other language is the only aim of the Language translation Systems. This
The translation
translation helpsofinone language to
socialization the other
means language
people is the onlywith
can communicate aim each
of theother
Language translation
in an easy way. AsSystems.
human This
is a
translation helps in socialization means people can communicate with each other in an easy way. As human is a

* Corresponding author. Tel.: +0-000-000-0000 ; fax: +0-000-000-0000 .

E-mail [email protected]
address:author.
* Corresponding Tel.: +0-000-000-0000 ; fax: +0-000-000-0000 .
E-mail address: [email protected]

1877-0509© 2019 The Authors. Published by Elsevier B.V.. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
1877-0509© 2019 The Authors. Published by Elsevier B.V.. This is an open access article under the CC BY-NC-ND license
Peer-review under responsibility of the scientific committee of the International Conference on Computational Intelligence and Data Science
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
(ICCIDS 2019)
Peer-review under responsibility of the scientific committee of the International Conference on Computational Intelligence and Data Science
(ICCIDS 2019)

1877-0509 © 2020 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the International Conference on Computational Intelligence and Data
Science (ICCIDS 2019).
10.1016/j.procs.2020.03.306
 Muskaan Singh et al. / Procedia Computer Science 167 (2020) 2534–2544 2535
2 Muskaan Singh/ Procedia Computer Science 00 (2019) 000–000

social entity, and from the ancient time they love to communicate with each other and live together. They
communicate with each other by using different means of communication and exchange their
data/information/thoughts with each other. Initially Sign Language is also the means of communication and is used
to exchange thoughts with each other but now in this era of technology there are number of effective sources that are
used for communication purpose. There are more than thousands of languages used all over the world even in India
due to its diverse nature and culture, different languages are used. In the local areas of the India, they have used their
own language of communication and Most of the literature is also available in the local languages. In India, there are
number of languages are used for the purpose of communication. Sanskrit is the mother of almost all Indian
languages and almost every language is generated from Sanskrit Language. As Sanskrit is very ancient language so
the main requirements in Sanskrit domain is to translate the life-transforming stories (epics), Vedas etc. to make
them available in other languages, for public at large. It is observed from various surveys that Sanskrit as a source or
target language is in developing stages and a major issue which arises in implementation of Sanskrit based MTS is
the approach used for developing it. There are number of historical and ancient granths which are written in Sanskrit
language, Bhagvad Gita is one of them and is very valuable in Indian Hindu culture.

Bhagvad Gita (BG) – the song of the lord included in Bhishma Parva (chapter 23 - 40) which is sixth book out of 18
books of Indian Epic Mahabharata [29]. There are 18 chapters and 700 slokas (verses) in BG. In BG, there are
number of combinations of different Hindu thoughts in regards to mystical bhakti, dharma and different yogic path
to find moksha. It is an essential content which abridges the Upanishadic lessons and is remarked upon and
translated by different schools of Indian theories. Being a vital content, it has been converted into every single real
dialect of the world, and furthermore remarked variously. In this way annotators in uncertainty can generally allude
to these critiques for right explanation. This sacred text being lucid and finish in itself, can be utilized for larger
amount examination, for example, talk investigation, point identification, anaphora goals, etc. So, it is an important
issue to convert this content and use it not only for research purpose but to save the history and culture of India. To
make this conversion automated, the field of Natural Language Processing (NLP) may be used. A lot of researchers
also worked on this to automate the translation systems and most of them selects NLP for the development of the
translators for the conversion of one language to other. NLP is defined as a mechanism which helps in the
understanding of the semantics and grammar of a language and accurate interpretation of the language [3]. NLP is
concerned with the development of models that automate work to process a language and help to make its use for
communication between humans. The basic model of the language processing system as shown in fig 1, requires
some knowledge to deal with the processing of the language and it also require some knowledge about language
along with its grammar [3]. Machine translation systems (MTS) are the systems which are totally based on NLP. In
today’s life, Machine translation is one of the applications of computers used to translate one natural language text
into the other language and it is an application of NLP. The term Machine Translation System called as MTS also
defined as the “translation from one natural language (Source Language -SL) to another language (Target Language
-TL) using computerized systems with or without human assistance”. These systems provide the translation solution
without any human interference or assistance. The main requirement of these systems is dictionaries and grammar
for those particular languages.

Natural Natural
Target
Language Language
Source Understand Generated
Language
Language

Fig 1: NLP Processing

1.1 Overview of MTS

With the advancement in the technology, the use of NLP is increased for the solution of different tasks. Now-a-days
2536 Muskaan Singh et al. / Procedia Computer Science 167 (2020) 2534–2544
Muskaan Singh / Procedia Computer Science 00 (2019) 000–000 3

Machine translation is one of the important tasks of NLPwhich is on high demand with the qualitative and efficient
results. These qualitative results benefitted into various fields [3] in various terms as discussed below:

 Searching and Extraction of Data: In the Era of technology, Internet becomes the source of information
for any field and for any one. So, the webpages where content is available might contains different
languages so, translation is required to enable the search and to extract data and remove language barriers.
 Translation of Technical Documents: The technical documents of different types like patent, research
articles, manuals and others professional documents which are used in various countries need to be
translated so everyone can have access to that data.
 Translation of Documents and Speeches:Some of documents and speeches which needs to float in
multiple countries need translation in multiple languages. For instance, In European Union (EU) they used
23 different official language because each member of European parliament speaks his/her own language,
because of this translation is required for each of these languages. The other important example is the
translation of the documents of United Nation is 6 different languages.
 Translation of Broadcasted Information:the media is one of the important sources of information where
information is broadcasted continuously and this public information is flowing all over the world so,
translation is required. There are many platforms that deals with the information transfer and these are
Radios, blogs, Newsfeeds, Television, and many more.
 And many more.

As discussed that the qualitative and efficient MTSs becomes the necessity of today, but the current technology
which develops the MTSs is not up-to-the-mark means not fulfil the expectations to deal with all the defined areas
but its continuous improvement originates the automated translation systems that are used for extracting information
or data.

1.2 Approaches used in MTS

In MTS, Source Language is first analyses and its internal representation is prepared accordingly and then it is
manipulated for transformation of it into its target language from which target language is generated. These
transformations don’t change the meaning of the input text throughout the whole process while translation. It may be
bilingual or multilingual. If translation is between the two language then it is known as bilingual systems but it is
between number of language then it is called as multilingual system [21]. The process of MTS may be defined in
two steps:

Step 1: Decoding: where meaning of the source text is decoded

Step 2: Re-encoding: where decoded text is re-encoded in target language.

Decoding is the process where source text is analyzed along with its all features so that translator will get the
accurate meaning of the source text. This analysis requires deep knowledge of the source language and other
parameters like, grammar, semantics, syntax etc. and similarly for re-encoding.

There are Number of MTSs which developed for the translations of some common languages like, English, Chinese,
Russian, Spanish, Japanese, Hindi and many other Indian and Foreign Languages. To develop these systems,
different approaches of MTSs has been followed and these are: (a) Rule Based Approach, (b) Dictionary based
Approach, (c) Corpus based Approach, (d) Knowledge Based Approach and (e) Hybrid Approaches as described in
Fig2.
 Muskaan Singh et al. / Procedia Computer Science 167 (2020) 2534–2544 2537
4 Muskaan Singh/ Procedia Computer Science 00 (2019) 000–000

•Based on Dictionary Entries

•Translation done word by word
Dictionary Based
•Morphological analysis may be used
Approach
•Least Sophisticated Approach
•Suitable for Long list of phrases
•Parsing based Approach
•Translation done sentence by sentence
Rule Based Approach •Morphological, Syntactic, and Semantic Analysis is done
•Based on Intermediate represntation
•Three Types: Direct, Transfer based and Interlingua
•Pre-translated parallel text is the main requirement of this
approach
Corpus Based •fully automatic
Approach •less tedious than rule based
•Availability of the corpora is one of the main difficulty
•Three Types: statistical, context-based and example-based
Knowledge Based
•use a machine readable dictionary, thesaurus or ontology
Approach
•combination of different approaches
Hybrid Approach •benefits of the combined approaches are included in this
approach
Fig 2: Approaches of MTS

1.3 Proposed Work

A new corpus-based machine translation system for Sanskrit to Hindi translation is proposed. There are very few
works done on the translators which translated Sanskrit to Hindi Language so need to develop an efficient translator
for the same. India is rich in culture and there are many Vedas and other holy books which were written in Sanskrit
like, Bhagavad Gita. For this work, Bhagavad Gita(BG) is selected as an Input Data Set and proposed a translator
which automatically translate it into Hindi Language.

The next sections of this paper describe some of the existing MTS for Sanskrit Language and the detail discussion of
this proposed MTS along with its simulation and results.

2. Existing Sanskrit based Translation System

Different researchers done work on Machine Translation Systems for Sanskrit Language. The discussion on some of
the proposed systems is given in this section. Tapaswi et al. [1] proposed a parsing technique named as Lexical
Functional Grammar (LFG) for Sanskrit text. LFG works on two basic types of syntactic representation: (a)
constituent structure, and (b) functional structure. They used LFG because translation was from Sanskrit to English
and both these language representations is different. For instance, English is Subject-Verb-Object (SVO) and
Sanskrit is Subject-Object-Verb (SOV). Their main aim to develop a parsing technique and testing of this is done on
simple sentences. The other work on lexical analysis is done by Tapaswi and Jain [2] for Sanskrit sentences. In this
work, morphological analysis was also added with lexical analysis. They designed their own rule format and stored
all the rules in the files with the names of its starting letter like ‘himalaya’ is stored in ‘h.txt’. Here the root word and
meaning of it is identified with the help of lexical analyzer.

Barkade et al. [3] proposed MTS for the translation of English to Sanskrit Language where they divided work into 4
modules: Lexical Parser, Semantic Mapper, ITranslator, and Composer and discussed first two modules in this
paper. They designed their own lexical parser for POS tag information and its dependency. Three different rules are
generated by this parser and are: Equality Rule, Synonym Rule, and Antonym Rule. After parsing when tokens were
2538 Muskaan Singh et al. / Procedia Computer Science 167 (2020) 2534–2544
Muskaan Singh / Procedia Computer Science 00 (2019) 000–000 5

generated and using dependency when relation between token is found then tree was generated and mapping done
between English and Sanskrit Sentence.EtranS- MTS for the translation of English to Sanskrit Language proposed
by Bahadur et al. [4] to improve quality of translation. They developed a software using .Net Framework and MS-
Access 2007. This software has two modules: (a) Parse Module where parsing was done means after analysis tokens
were generated and grammatical and syntax analysis was done. (b) Generator Module uses semantic information for
mapping and on the basis of mapping results were generated. They tested their proposed system for small, large and
extra-large sentences and achieved 99% correctness for small and 90% for extra large sentences.
Mishra and Mishra [11] compared different MTS systems and analyzed the performance of Example based system
for English to Sanskrit translation. They used ENGCC parser for English and for Sanskrit parser is used which was
developed by Gerard Huet(http://sanskrit.inria.fr/) and dictionary is used
from(www.dicts.info/dictionary.php?l1=English&l2=Sanskrit) site. They analyzed that the EBMT is used for scarce
online resource’s data and worked properly for results. The other work done by Mishra [24] where they proposed
English to Sanskrit based MTS using rule-based approach. Pandey and Jha [18] analyses error for Sanskrit to hindi
MTS that uses statistical approach. They build corpus and trained using MTHub platform. The error report
generated by MTHub System and during training of two phases BLEU score was calculated. In the first phase,
10000 long, complex and compound sentences were used where BLEU score was 39.17 and in second phase, 24000
bilingual and 25000 monolingual sentences were used and achieved 41.17 BLEU score.

3. Proposed Corpus Based Machine Translation System with Deep Neural Network

The main focus of this work is to propose a novel approach to develop corpus-based MTS with deep neural network
for translating Sanskrit to Hindi Text and here corpus-based MTS is developed which doesn’t need any rules or
dictionaries because they automatically learn about language from a large set of corpora. The corpus is improved
with the addition of trained data using deep neural network to learn all the expressions like phrasal and idiomatic
expressions easily. This work is done in three different phases named as (a) Data Analysis, (b) Data Processing and
(c) Target Generation where firstly data analysis is done and then data will transfer and generated into target
language and lastly final generation of target text using Deep Neural network. Here, deep learning provides better
training and accurately find the results.

Proposed Model
This proposed corpus-based model with deep neural network divides the work into three phases named as: (i) Data
Analysis, (ii) Data Processing and (iii) Target Generation. These phases further perform different operations to
generate target text as shown in figure below:
Data Tokenizatio
Collection n

Data Training
Data Parsing Target
Data Cleaning
Analysis Processin Generatio
Data Pre- g Splitting n
processing Testing

Data Pass
Visualizati toHidden
on Layer
Fig 3: Phases in Proposed MTS
Phase 1: Data Analysis
This step includes the collection the dataset containing millions of word meanings in a systematic way from the
different resources to make it useful for proposed model. From all sparse spreader datasheets make big dataset and
save it in a proper word meaning format. If dataset is not in correct format, mark it as redundant dataset. After
collecting all the dataset in a single csv, clean the dataset by removing all the redundant or unused words and extra
symbols present in dataset. Here the dataset is arranged in a proper structure to make it suitable for the proposed
model. After that, visualize the dataset to check the structure and the co-relation in the dataset. If there is no relation
found in dataset then pre-processing of the data is again performed from starting.
6 Muskaan Singh/ Procedia Computer Science 00 (2019) 000–000
Muskaan Singh et al. / Procedia Computer Science 167 (2020) 2534–2544 2539

Phase 2: Data Processing

Tokenization is the chopping of data into words. In this step store all the dataset in matrix format and then Tokenize
the data by splitting all the sentences into words and label the data into numeric form. As this data is in Sanskrit and
Hindi language, encode it normally as done for English words So here use Count vectorise. Count-Vectorise will
convert all type of data to numeric form easily so use it for labelling the data. After tokenization data is analysed
through lexical and semantic analysis. Here arrange the data in a proper grammatical structure with respect to the
Sanskrit rules. To overcome the problem of overfitting and under fitting, the data is split in two parts out of which
one is used for the model development for predictive analysis and the other one is used for performance analysis.
Divide the data into two parts for training and testing purpose.

Table 1: Example of Data Processing

Input
Sentence
Tokenization

Semantic
Analysis

Parsing

Morphology

Target
Sentence

Phase 3: Target Text Generation

In the proposed system, training is performed using deep neural network. Here input is passed neural network with a
huge number of hidden layers that layers will be added automatically through auto-tuning. The auto-tuning has
2540 Muskaan Singh et al. / Procedia Computer Science 167 (2020) 2534–2544
Muskaan Singh / Procedia Computer Science 00 (2019) 000–000 7

helped to make this model better than all other previous models. Now, system will automatically calculate required
number of input layers which makes proposed system accurate and then it is trained using large number of hidden
layers at a very high speed which makes it much better.

So after processing, now pass the training set into model. Here, Input is passed through first layer of neural networks
and number of neurons. Secondly, applied activation functions and gives probability of output. At the end, it passes
output of first NN to second layer NN. Input layer is the layer which interacts with hidden layers. The accuracy of
the system is depended on the number of input layers passed and the number of times it interacts with the hidden
layers. More the interaction between the input layers and hidden layers more is the accuracy attained and vice versa.
Now, first of all, normalize the batch and again apply activation function. Passes output of second NN to third layer
NN. Again, normalizing the batch and apply activation function. Then Passes output of Third NN to fourth layer
NN. Repeat this process up to seventh layer NN. After addition of layers, divide and merge the datasets according as
per the requirement. Normalize the batch again. Again, applied activation functions and gives probability of output.
Again, divide and merge the data as per requirement neurons are passed in dense step. Then output is generated at
the end that exhibits the accuracy of 99.97%.

4. Experiment Analysis

In this section, details related to experimental analysis has been discussed for corpus-based machine translation
system with deep neural network.

4.1 Simulation Setup

In the proposed work, Keras sequential model is used to process the data. Proposed model is processed through
highly configured core GPU with 32 GB of RAM to achieve a high throughput speed approximately 2500 words per
second. This speed is not possible for normal systems because in this one epoch will take approximately two hours
to run. Sothe use of highly configured GPU along with NVIDIA Geforce GTX 980 GPU is preferred here.

4.2 Dataset Used

For the proposed model huge amount of dataset is collected which consists of Sanskrit to Hindi word meanings as
shown in Table 2. Along with the word meanings data it also consists of various rules of Sanskrit as well as Hindi.
Thus, redundant data from that dataset has been removed and arranged in proper word meaning format to pass it
through the model. Also, collect the huge dataset in a single csv file which contains around millions of sentences.

Table 2: Dataset
Dataset #Sentences #Words Vocabulary
Source Target Source Target
Corpus: Open MT Sanskrit-English
Training 145,34,215 131575835 123425654 355.465 124.278
Development 192679 172799 122645 NA NA
Test 12698 77322 26273 NA NA
Corpus: Open BTEC English-Hindi
Training 123643124 11425429 11753927 428.672 111.249
Development 12679 122769 87286 NA NA
Test 10684 67827 39207 NA NA

The collected data was non-uniform and unformatted data. Therefore, various ways are needed to make it suitable
for proposed model. Single csv file is processed twice for rechecking to remove redundant words. If any redundant
 Muskaan Singh et al. / Procedia Computer Science 167 (2020) 2534–2544 2541
8 Muskaan Singh/ Procedia Computer Science 00 (2019) 000–000

words are found it removes them from there, otherwise it does encode the dataset.

4.3 Corpus Based Results

BG has 700 sections. Greater part of these refrains (645) are created in a measurement called anustup and the rest of
the sections are in indravajra, upendravajra and upajati meters. A series of characters isolated by spaces may
compare to at least one words. The aggregate number of string successions isolated by spaces is 6426 adding up to
9.18 words per section. In the wake of part these strings into words, there were 8884 words 13 at the end of the day
after division there was around 13.82% expansion in the words. Out of 8884, 1413 words were observed to be mixes
adding up to 15.9%. The some of the resultant Hindi text generated by using this proposed Model is as shown in
table 3.

Table 3: Sanskrit Sentences and their corresponding Hindi Sentences generated by using proposed model
Sanskrit Sentences Hindi Sentences
1 xk.Mhoa L=alrsgLrkÙoDpSoifjng~;rsA gkFk ls x.Mho /kuq’kfxjjgkgSvkSjRopkHkhcgqrtjjgh g
u p ‘kd~uksE;oLFkkrqaHkzerho p SrFkkesjkeuHkzfer&lkgksjgkgS bl fy;seSa [kMkjgusdksleFkZ ugh gqaA
esaeu%AA
2 fufeÙkkfu p i’;kfefoijhrkfuds’koA gsds’ko! eSy{k.kksa dk Hkhfoifjrgh ns[k jgkgqWrFkk ;q) essaLotuleqnk;
u pJs;ks·uqi’;kfegRokLotuekgosAA dksekjdjdY;k.kHkh ugh ns[krkA
3 udkM~{ksfot;ad`”.k u p jkT;alq[kkfupA gsd`”.k! eS u rksfot; pkgrkgqWvkSj u jkT; lq[kksadksghAgsxksfoUn! gesa
fdauksjkT;suxksfoUnfdaHkksxSthZforsu ,slsjkT; ls D;kiz;kstugSvFkok ,slsHkksxks ls vkSj thou ls HkhD;kykHkgSA
okAA
4 ;s”kkeFksZdkM~+f{krauksjkT;aHkksxk% gesaftudsfy;sjkT;]HkksxvkSjlq[kkfnvHkh’VgSa] osgh ;s lc /kuvkSj thou dh
lq[kkfupA vk’kkdksR;kxdj ;q) es [kMsgSaA
rbes·ofLFkrk ;q)s izk.kkaLR;DRok
/kukfupAA
5 vkpk;kZ% firj% iq=kLrFkSo p firkegk%A xq:turkm&pkps] yMdsvkSjmlhizdkjnkns] ekes] llqj] iks+= lkysrFkkvkSjHkh
ekrqyk% Üo’kqjk% ikS=k% ‘;kyk% lEcU/khyksxgSA
lEcfU/kuLrFkkAA
6 ,rkUugUrqfePNkfe ?urks·fi e/kqlwnuA gs e/kqlwnu !eq>s ekusijHkhvFkokrhuksayksdksa ds jkT; ds
vfi =SyksD;kjkT;L; gsrks% fy;sHkheSabudksekjukughpkgrk! gQji`Foh ds fy;srksdgukghD;kgSA
fdauqeghd`rsAA
7 fugR; /kkrZjk”VªkUu% dkizhfr% gstuknZu ! /k`rjk”Viq+=ksadksekjdjgesaD;kizlUurkgksxh !
L;kTtuknZuA buvkrrkf;;ksadksekjdjrksgesaikighyxsxkA
ikiesokJ;snLekUgRoSrkukrrkf;u%AA
8 rLekUukgkZo;agUrqa vr ,ogsek/ko ! viusghckU/ko /k`rjk”Vª ds iq+=ksadksekjus ds fy;sge ;ksX;
/kkrZjk”VªkULockU/koku~A ugh gS%] D;ksafdviusghdqVqEcdksekjdjgedSlslq[khgksxsaA
LotuafgdFkagRoklqf[ku% L;keek/koAA
9 ;|I;srs u i’;fUryksHkksigrpsrl%A ;|fi yksHk ls Hkz”Vfprgq, ;s yksxdqy ds uk’k ls mRiUunks”kdksvkSjfe+=ksa
dqy{k;d`ranks”kafe=nzksgs p ikrde~AA ds fojks/k djusesikidks ugh ns[krs] rksHkhgstuknZu! dqy ds uk’k ds
dFka u Ks;eLekfHk% mRiUunks”kdkstkuusokysgeyksxksbliki ls gVus ds fy;sD;ks ugh
ikiknLekfUuofrZrqe~A fopkjdjukpkfg;sA
dqy{k;d`ranks”kaizi’;fHntZuknZuAA
10 dqy{k;s iz.k’;fUrdqy/kekZ% lukruk%A dqy ds uk’k ls lukrudqy&/keZu”VgkstkrsagSa] /keZ ds
/kesZu”Vsdqyad`RLue/keksZ·fHkHkoR;q uk’kgkstkusijlEiw.kZdqy es ikiHkhcgqrQSytkrkgSA
rAA
11 v/kekZfHkHkokRd`”.kiznq”;fUrdqyfL=;% gsd`”.k! iki ds vf/kd c<+ tkus ls dqy dh
A fL=;kavR;Urnwf”krgkstkrhgSvkSjok”.ksZ;! fL+=;ksa ds
L=h”kqnq”Vklq ok”.ksZ; nwf”krgkstkusijo.kZladjmRiUugksrkgSA
tk;rso.kZlM+dj%AA
2542 Muskaan Singh et al. / Procedia Computer Science 167 (2020) 2534–2544
Muskaan Singh / Procedia Computer Science 00 (2019) 000–000 9

12 lM+djksujdk;Sodqy?ukukadqyL; pA o.kZladjdy?kkfr;ksadksdqydksujdystkus ds fy;sgksrkgS A

irfUrfirjksás”kkayqIrfi.MksndfØ;k% AA yqIrgqbZfi.MvkSj ty dh fØ;kokysvFkkZr~ Jk) vkSjriz.koafprbldsfirjyksxHkh
v/kksxfrdksizkIrgksrsgSA
13 nks”kSjsrS% blo.kZladjdkjdnks”kksa ls dqy?kkfr;ksa ds
dqy?ukukao.kZlM+djdkjdS%A lukrudqy&/keZvkSjtkfr&/keZu”V
mRlk|Urstkfr/kekZ% dqy/kekZÜp gkstkrsgSaA
‘kkÜork%AA
14 mRl=dqy/kekZ.kkaeuq”;k.kkatuknZuA gstukjnZu !ftudkdqy /keZu”Vgksx;kgS] ,slseuq”;ksadkvfuf’pedkyrd
ujds·fu;raoklksHkorhR;uq’kqJqeAA ukjhesaoklgksrkgS] ,slkgelqursvk;sgSaA
15 vgkscregRikiadqrZaO;oflrko;e~A gka! ‘kksd! geyksxcqf)ekugksdjHkhegku~ ikidjusdksrS;kjgksx;sgS] tks
;nzkT;lq[kyksHksugUrqaLotueq|rkAA jkT; vkSjlq[k ds ykHk ls Lotuksadksekjus ds fy;sm|rgksx;sgSA
16 ;fnekeizrhdkje’kL=a ‘kL=ik.k;%A ;fneq> ‘kjfgr ,aolkeuk u djusokysdksgkFkesafy;sgq, /k`rjk”Vª ds iq+=
/kkrZjk”Vªkj.ksgU;qLrUes j.kesaekjMkysrksogekjukHkhesjsfy;svf/kdDy;k.kdkjdgksxkA
{kserjaHkosr~AA
17 ,oesqDRoktqZu% lM~[;s lat; cksys&j.kHkwfeeas ‘kksd ds mf)Xu euokysvtqZublizdkjdgdjok.k
jFkksiLFkmikfo’kr~A lfgr /kuq”kdksR;kxjdjFk ds fiNysHkkxescSBx;sA
fol`T; l’kjapkia ‘kksdlafoXuekul%AA

4.4 Performance Metrics

To evaluate the performance of this proposed corpus based approach with deep neural network, BLEU Score and
Word Error Rate is calculated in this work.
 BLEU Score

Bleu score is an important metric used for calculating the accuracy of translated sentences as compared to the human
generated reference translations. It is not good for shorter translations but it provides accurate results for longer
sentences. Normally Bleu Score values lies between 0 and 1, simply multiplying it to 100, its percentage can be
calculated. It is observed that higher the bleu score value, model is more accurate. Formula of Bleu Score is as
follows:
�������������
𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 𝐵 𝐵𝐵𝐵 �1, � (∏���� 𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝� ) ……..(i)
������������

BLEU Score

0.5

0
RBMT Corpus based MTS
with Deep Neural
Network
Fig 4: BLEU Score of Rule Based and Proposed MTS
Fig 4 shows that this proposed machine translation system achieves better BLEU Score as compare to rule based
machine translation systems. As per results when training of corpus is done with deep neural network then the
performance of the corpus-based MTS is 24% better than RBMT.

 Word Error Rate

It is a metric used to calculate the error rate by comparing machine translated output with the human translated
output. It is assumed that if less the WER, better the model will be
 Muskaan Singh et al. / Procedia Computer Science 167 (2020) 2534–2544 2543
10 Muskaan Singh/ Procedia Computer Science 00 (2019) 000–000

���������������������������������
𝑊𝑊𝑊𝑊𝑊𝑊 𝑊 ………(ii)
�������������
Here substitution means replacement of one word with another. Insertion means addition of words and deletion
means dropping of words.

Word Error Rate

1
0.5
0
RBMT Corpus based MTS with
Deep Neural Network
Fig 5: WER of Rule Based and Proposed MTS

Fig 5 shows that the proposed machine translation system has less word error rate which means its performance of
the proposed MTS is better than Rule based Machine Translation System by 39.6%.

5. CONCLUSION

Sanskrit to Hindi Translation is one of the most challenging tasks. As a result model became complex and time
consuming. In this paper, to overcome existing problem, deep learning concept is used to train the data and for
model building. Tensor flow library is used to build a model. In this proposed work, Keras is used as a front end and
Tensor flow is used as a back end library. Performance evaluation show that proposed MTS system gives better
performance in terms of BLEU Score and WER.

REFERENCES
[1] N. Tapaswi, S. Jain, and V. Chourey, (2012) “Parsing Sanskrit Sentences using Lexical Functional Grammar”, International
Conference on Systems and Informatics, pp. 2636-2640.
[2] N. Tapaswi, and S. Jain, (2011) “Morphological and Lexical Analysis of the Sanskrit Sentences” MIT International
Journal of Computer Science & Information Technology, Vol. 1, No. 1, pp.28-31.
[3] V. M. Barkade, and P. R. Devale,(2010) “English to Sanskrit Machine Translation Semantic Mapper”, International Journal
of Sciemce and Technology, Vol. 2, Issue 10, pp. 5313-5318.
[4] P. Bahadur, A.k.jain, and D.s.chauhan,(2012) “EtranS- A Complete Framework for English To Sanskrit Machine
Translation,” International Journal of Advanced Computer Science and Applications, vol. 2, no. 1, pp. 7–13.
[5] U. Germann, M. Jahr, K. Knight, D. Marcu, and K. Yamada, (2004) “Fast and optimal decoding for machine
translation,” Artificial Intelligence, vol. 154, no. 1-2, pp. 127–143.
[6] T. Xiao, J. Zhu, and T. Liu, (2013) “Bagging and Boosting statistical machine translation systems,” Artificial Intelligence,
vol. 195, no. 6, pp. 496–527.
[7] R. U. U. and T. A. Faruquie, (2005) “An English-Hindi Statistical Machine Translation System,” Natural Language
Processing – IJCNLP 2004 Lecture Notes in Computer Science, vol. 5, no. 1, pp. 254–262.
[8] Y. Zhang,(2017) “Research on English machine translation system based on the internet,” International Journal of Speech
Technology, vol. 20, no. 4, pp. 1017–1022.
[9] C. Adak, (2014) “A bilingual machine translation system: English & Bengali,” 2014 First International Conference on
Automation, Control, Energy and Systems (ACES), pp. 1–4.
[10] L. Chirong, (2014) “Research and Implementation on Machine Translation System with Online Corpora Extraction
Technology,” Fifth International Conference on Intelligent Systems Design and Engineering Applications, pp. 759–763.
[11] V. Mishra, and R. B. Mishra, (2008) “Study of Example Based English to Sanskrit Machine Translation”, Polibits Journal,
Vol. 37, pp. 43-54.
[12] S. Kharb, H. Kumar, M. Kumar, and A. K. Chaturvedi, (2017) “Efficiency of a machine translation system,” International
conference of Electronics, Communication and Aerospace Technology (ICECA), pp. 140–148.
2544 Muskaan Singh et al. / Procedia Computer Science 167 (2020) 2534–2544
Muskaan Singh / Procedia Computer Science 00 (2019) 000–000 11

[13] J. Nair, K. A. Krishnan, and R. Deetha, (2016) “An efficient English to Hindi machine translation system using hybrid
mechanism,” International Conference on Advances in Computing, Communications and Informatics (ICACCI), pp. 2109–
2113.
[14] B. N. Raju and M. B. Raju, (2016) “Statistical Machine Translation System for Indian Languages,”IEEE 6th International
Conference on Advanced Computing (IACC), pp. 174–177.
[15] S. Saini and V. Sahula, (2015) “A Survey of Machine Translation Techniques and Systems for Indian Languages,” IEEE
International Conference on Computational Intelligence & Communication Technology, pp. 676–681.
[16] R. Östling, Y. Scherrer, J. Tiedemann, G. Tang, and T. Nieminen, (2017) “The Helsinki Neural Machine Translation
System,” Proceedings of the Second Conference on Machine Translation, pp. 338–347.
[17] D. Ze-Ya, Z. Han-Fen, Z. Quan, M. Jian-Ming, and C. Yu-Huan, (2009) “Automatic Machine Translation Evaluation
Based on Sentence Structure Information,”International Conference on Asian Language Processing, pp. 162–166.
[18] R. K. Pandey and G. N. Jha, (2016) “Error Analysis of SaHiT - A Statistical Sanskrit-Hindi Translator,” Procedia
Computer Science, vol. 96, no. 6, pp. 495–501.
[19] P. Goyal and R. M. K. Sinha, (2008) “Translation Divergence in English-Sanskrit-Hindi Language Pairs,” Lecture Notes in
Computer Science Sanskrit Computational Linguistics, vol. 8, no. 1, pp. 134–143.
[20] P. Shukla and A. Shukla, (2013) “A Framework of Translator From English Speech To Sanskrit Text ,” International
Journal of Emerging Technology and Advanced Engineering , vol. 3, no. 11, pp. 113–121.
[21] A. Godase and S. Govilkar, (2015) “Machine Translation Development for Indian Languages and its
Approaches,” International Journal on Natural Language Computing, vol. 4, no. 2, pp. 55–74, 2015.
[22] N. Sadana, (2017) “Comparison Of Sanskrit Machine Translation Systems,” International Journal of Advanced Research in
Computer Science, vol. 8, no. 8, pp. 223–225.
[23] J. K. and J. R., (2016) “Sanskrit Machine Translation Systems: A Comparative Analysis,” International Journal of
Computer Applications, vol. 136, no. 1, pp. 1–4.
[24] V. Mishra and R. Mishra, (2012) “English to Sanskrit machine translation system: a rule-based approach,” International
Journal of Advanced Intelligence Paradigms, vol. 4, no. 2, pp. 168–184.
[25] J. K. Raulji and J. R. Saini, (2017) “Generating Stopword List for Sanskrit Language,” 2017 IEEE 7th International
Advance Computing Conference (IACC), pp. 799–802.
[26] S. J, J. S, and D. K. R.N, (2014) “An Efficient Machine Translation System for English to Indian Languages Using Hybrid
Mechanism J,” International Journal of Engineering and Technology, vol. 6, no. 4, pp. 1909–1919.
[27] H.M. Parmar,(2015) “A Toolkit for Sanskrit Processing”.
[28] S. Saini and V. Sahula, (2015) “A Survey of Machine Translation Techniques and Systems for Indian Languages,” IEEE
International Conference on Computational Intelligence & Communication Technology, pp. 286–290.
[29] Shukla, P., Kulkarni, A., and Shukl, D., (2013) “Geeta: Gold Standard Annotated Data, Analysis and its Application” ,
ICON.