Archives of Computational Methods in Engineering

https://doi.org/10.1007/s11831-020-09449-7

ORIGINAL PAPER

Machine Translation Systems for Indian Languages: Review

of Modelling Techniques, Challenges, Open Issues and Future Research
Directions
Muskaan Singh1 · Ravinder Kumar1 · Inderveer Chana2

Received: 23 October 2019 / Accepted: 1 June 2020

© CIMNE, Barcelona, Spain 2020

Abstract
With the advancement in computer language technology in a multilingual country like India, numerous linguistics require
technology for translation. It aids in research for ancient languages like Sanskrit, Tamil, Telugu, Malayalam to be available
for society. Devising these languages require natural language processing, and machine translation is one of its essential
areas. It plays a significant role in breaking the language barrier and facilitating inter-lingual communication by translating
one language to another. As, with the advent of information technology, many documents and web pages are available in
local languages. There is a tremendous need to establish proper communication amongst the people belonging to discrete
backgrounds and cultures. This paper contributes in two ways: firstly, the review of different modelling techniques is per-
formed. It serves the developers with resources required for modelling different techniques such as corpus, domains, toolkits,
techniques, models, features and their evaluation measures. Secondly, a comparison of research work on different Indic
language pairs based on their modelling techniques have been performed. It influences the work on Sanskrit language to
be minimal despite holding an ancient scientific and comprehensive literature of India. It also contributes to linguistic and
technical challenges for processing Sanskrit language, open issues, and future research directions in this field.

1 Introduction information extraction, sentiment analysis, speech recogni-

Artificial intelligence (AI) aims to develop an intelligent sys-
tem examined by humans intuitively [1]. Natural language
processing (NLP) is one of many applications of AI. NLP is
an area of research and application involving computers to
understand the text in natural language. It builds a computa-
tional model for its analysis and generation. It involves tech-
nological, cognitive and linguistic motivation for developing
intelligent computer systems such as machine translation,
* Muskaan Singh
muskaan_singh@thapar.edu
Ravinder Kumar
ravinder@thapar.edu
Inderveer Chana
inderveer@thapar.edu
1 formalisms used in logic, frames and semantic networks.
Language Engineering and Machine Learning Research Lab,
Thapar Institute of Engineering and Technology (TIET),
Patiala, Punjab, India
2
CSED, Thapar Institute of Engineering and Technology
(TIET), Patiala, Punjab, India
sation, text classification, etc. Machine translation (MT) is
one of key applications of NLP [2, 3].
1.1 Machine Translation
MT is a process of translating source language to target lan-
guage using a computerised system. Human translators or
editors can be involved in the process of MT, although mini-
mal human aid is the goal of MT. The field of man–machine
interaction involves the processing of natural language.
Some of the technologies that aid in the development of
MT have been listed below:
• Computational Linguistics It covers word formations and
ordering, analysis of meaning and other communications
aspects.
• Knowledge Representation It is an area that deals with
• Semantic Network It provides linkage through relation-
ships of concept collections.

13
Vol.:(0123456789)
 M. Singh et al.

• Machine Learning It is a field where the machine learns
from different machine learning models for the accession
of new knowledge from data.
• Search Algorithms It assists in finding a solution to the
problem and not getting stuck in an infinite loop.
A translation provided by the Machine Translation System
(MTS) involves both syntactic and semantic aspects of lan-
guage to be covered to provide the correct version of transla-
tion [4]. MT is considered a challenging task as it includes
immense statistical models that have been created by using
complex semantic information. It is a sub-field of computa-
tional linguistics that examines the utilisation of program-
ming to disambiguate sentences at the discourse level. At
the basic level, MT performs substitution of words from one
language to another using a dictionary. Substitution of words
does not provide a sound translation of a text because rec-
ognition of whole phrases and their closest counterparts in
the target language is required. Improving translation quality
involves corpus improvement, human involvement, efficient
approach, handling contrasts in etymology, typology and the
segregation of anomalies. MTS has its applicability in enter-
prises, professional translation, casual/home use, bilingual
communication and bilingual dictionaries. It may act as a
benchmark for human professional translators. The uses of
MTS are categorized into dissemination, assimilation, inter-
action and translation aids [5, 6]. The properties of MTS
have been depicted in the Table 1.
In any case, even with these advantages as referred to
in the Table 1, MT still hasn’t surpassed human translation
in terms of translation accuracy and fluency of translation.
Although minimal human intervention is the goal of MT
which has been approached by recent techniques of deep
learning. In the prevailing scenario, MT output still needs
a human translator to edit and post-process its output for
effective translation. As web opens up to the more exten-
sive multilingual and worldwide group, innovative work on
MT keeps on developing at a quick rate. A couple of MTS
are accessible commercially but minimal work for Indian
languages has been performed as compared to European
Languages. MTS utilises bilingual information in the form
of corpora and other linguistic resources for processing the
language and machine learning models. Building the MTS
requires skill along with immense knowledge in punctua-
tion, linguistic structure and semantics in the source and
target language. Human and MT each have their share of
difficulties for developing MTS. It also requires to choose
the appropriate modelling technique which is based on
different criteria as listed.
• Resource availability (parallel corpus, monolingual cor-
pus, human resources, lexical resources and technical
resources).
• Background of the developer (linguist prefer rule-based
technique, translators prefer example-based technique,
computer-scientist prefer interlingua-based technique,
statistician prefer statistical-based technique and math-
ematician prefer neural-based technique).
• Goal or purpose of MTS (assimilation, dissemination,
one or more language pair, general or specific domain).

Table 1  Properties of MT system
Properties Description

Quick translation MT system enables you to save your time while translating large texts
Enhanced timeline A key benefit of MT is speed. It is significantly faster than human translators. It is estimated that an average human
translator could translate around 2000 words per day. Also, multiple translators can be assigned to a given project
to increase the output. Still, it lacks in comparison to MT as MT can generate thousands of words every minute
Confidentiality Many people use MTS to translate their private emails because no one would agree to give his private corre-
spondence to a translator whom he does not know. Also, no one would like to entrust financial documents to an
unknown person
Cost The cost of MT is also significantly lower than that of human translation. There are incredibly advanced MT
platforms that are accessible for free. Google translation is becoming more accurate each year and users can also
enter an unlimited amount of contents for translation for free
Universality Usually, a professional translator has specialisation in a particular field, but the MTS can translate any text about
any particular field. For the translation of specialised terminology, you have to switch to a corresponding setting
Consistency MT allows systems to memorise key terms and phrases used within a given area. This leads to increase in consist-
ency across the entire text. If human translators are used for translating a text, their translations may change
slightly depending on many factors whereas a MT output is consistent
Online translation and The advantage of online translation services is evident. Online translation services are at hand, and you can
translation of web page translate information more quickly. Furthermore, you can translate any web page content and query of the search
content engine by the use of MTS

13
Machine Translation Systems for Indian Languages: Review of Modelling Techniques, Challenges,…
1.2 Research Problem
The process of communication comprises of coding and
decoding. The speaker encodes his thoughts in a language
string and listener decodes it. The challenge in this com-
munication process arises with the restriction that lan-
guage inflicts both at word and sentence level. The word
choices for any language are many but finite, causing a
challenge in the communication of proper selection of
word. These words are then combined by language impos-
ing grammatical rules restraining the speaker to encode
his thoughts. Thus the encoding process for the speaker
has specific challenges for communicating his thought
to the listener. Therefore, approximation methods are
approached in the real world scenario. Thus thoughts
of the speaker are restricted by language constraints
and the encoded information is only an approximation
of thoughts. The computer is an information processing
device and used since its invention in the field of NLP.
Since 1950’s efforts are being carried out to build auto-
matic MTS. Earlier the difficulties were not recognized
and the researchers were enthusiastic for the advancement
in this field. This lead to the advancement in linguistic
tools, statistical, deep learning techniques and computer
hardware. As a consequence, this field developed with
better open research theories, data and tools for process-
ing the language [7].
Sanskrit is the ancient language of India with an enor-
mous collection of literature in different domains such
as medicine, poetics, mathematical logic, astronomy,
literature, technology, philosophy and dramatics. With a
considerable literature of its own, it has lived as a spoken
language for almost 1000 years. Until recent times it was
a means of communication for scholarly and discourse
communications. This produced the perpetual generation
of literature in Sanskrit for different domains for almost 2
millennia. The corpus size of Sanskrit is 100 times more
than those of Latin and Greek. Besides literary work,
there is a largely grammatical and philosophical tradition
that has lasted to live with undiminished strength until the
present time [8]. Nevertheless, in the last two centuries,
the situation has completely changed. Western learning
systems have substituted traditional learning methods. As
an outcome, the vast reserves of knowledge in Sanskrit
text is unavailable to Indian scholars [9]. Therefore there
is a need for an efficient MTS that can easily translate
Sanskrit to Hindi so that the vast scholarly and other
traditional knowledge is readily available to the current
generation. This was our main motivation for carrying
out review work addressing Sanskrit language and to lead
this research area for further exploration by deriving chal-
lenges and opportunities in this field.
1.3 Brief Overview of Recent Works
This section examines the various review paper available for
MTS. It distinguishes the generic survey with the Sanskrit
specific survey. Although several reviews have been carried
out for MTS addressing Indic Languages, only two surveys
have been found for processing Sanskrit as the source, target
and key supporting languages [10, 11]. Most of the previ-
ous works for MT were generic and gives a brief descrip-
tion of MT approaches along with major MTS develop-
ment in India for English to Indian languages and Indian
to Indian languages [12]. The survey conducted by [13]
also describe briefly the MTS for Indian languages along
with their features, domain and limitations. Another survey
by [14] reviewed only important MTS focussing on their
methodology. Naskar and Bandyopadhyay [15] and Dwivedi
and Sukhadeve [16] discusses details of MTS developed for
English-Indian languages and Indian to Indian languages as
a field or Web service.
The review studies specific to Sanskrit language [10, 11]
describes briefly MT approaches and MTS for the Sanskrit
language. Many other aspects such as challenges, lexical and
technical resource requirements prevailing to the Sanskrit
language need to be considered before developing MTS. The
resources required for developing MTS such as appropriate
modelling technique and their specific linguistic and tech-
nical requirements such as corpus, domains, toolkits, tech-
niques, models, features. The novelty of the current survey
focuses on the resources of different MT modelling tech-
niques for Indian language research. The study of model-
ling techniques was compared for ancient languages such as
Punjabi, Bengali, Marathi, Telugu, Tamil, Assamese, Urdu,
Malayalam, Gujarati, Sanskrit, Kannada, Dogri, Sinhala
and Devnagari and from the synthesis, it was concluded that
work for the Sanskrit language was minimal despite such
rich literature. Therefore this work, contribute to research
for the Sanskrit language with all the resources, challenges,
future research directions in this field. Taking care of recent
advances in deep learning, it is highly required to analyse
the research in this dimension. Therefore, the current new
paradigm in MT i.e Neural Machine Translation has been
reviewed along with its modelling technique, corpus, tools,
domains, technique, features and evaluation measures. The
technical challenges for applying this modelling technique
for the Sanskrit language has also been covered in Sect. 3
along with linguistic challenges. The open issues and future
research direction in this field have also been highlighted.
1.4 Goal of Paper
This section discusses the goal of the paper with the help of
a few highlights contributing to the paper.
13

1. To identify gaps in the existing research work for devis-
ing an effective MT technique.
2. The MTS should easily translate Sanskrit to Hindi so
that the vast scholarly and the other traditional knowl-
edge is readily available to the current generation.
3. The review of different modelling techniques with a per-
spective of resources has been presented in the paper. It
serves the developers with resources required for model-
ling different techniques such as corpus, domains, tool-
kits, techniques, models, features and their evaluation
measures.
4. A comparison of research work on different Indic lan-
guage pairs based on modelling techniques has been
performed. The survey of modelling techniques was
compared for ancient languages such as Punjabi, Ben-
gali, Marathi, Telugu, Tamil, Assamese, Urdu, Malay-
alam, Gujarati, Sanskrit, Kannada, Dogri, Sinhala and
Devnagari and from the synthesis, it was concluded that
work for the Sanskrit language was minimal despite such
rich literature. Therefore this paper contributes to the
research of MTS for processing the Sanskrit language.
5. The review also covers the dimension of recent advances
of deep learning in MT paradigm i.e Neural Machine
Translation.
6. The critical challenges, technical as well as linguistic
that are likely to be faced in building MTS for the San-
skrit language have been reported in depth.
7. Open issues and future research direction in this field
have also been discussed in this paper.
1.5 Organization of Paper
The remaining paper is organised into following different
sections. Section 2 describes the background of MTS as a
brief introduction to MT and its approaches. Section 3 con-
tains challenges pertaining to Sanskrit languages. It covers
different linguistic and technical challenges. Section 4 con-
tains a comparison of various modelling techniques based
on the resource requirement. Section 5 compares MTS for
different ancient languages while section 6 compares MTS
Fig. 1  Classification of transla-
tion modelling
13
M. Singh et al.
only for processing the Sanskrit language. This paper further
contributes to Open Issues in Sect. 7, future research direc-
tions in Sect. 8 and finally conclusion in Sect. 9.
2 Background: Machine Translation
Modelling Techniques
This section describes the pre-eminent MTS modelling tech-
niques in brief. These techniques have been classified on
the basis of engineering involved for developing it either
manually or mechanically. The techniques have been majorly
categorised into human-engineered, machine-engineered and
combination of both i.e hybrid. The human-engineered mod-
elling technique is rule-based and machine-engineered is
corpus-based. These techniques have been further classified
as shown in the Fig. 1 and described in further subsections.
2.1 Human‑Engineered Translation
This modelling technique involves rule-based MTS which
can be build using dictionary-based, transfer or interlingua
approach. It involves more human intervention as all the
modules of these system requires human insights.
2.1.1 Rule‑Based Machine Translation (RBMT) Modelling
The RBMT modelling technique is one of the oldest and
is still being used for less-resourced languages [17]. This
technique depends on the linguistic features of the source
and target language. The linguistic information along with
grammatical properties (morphological, semantic and syn-
tax) is acquired from lexical resources such as bilingual,
unilingual or multilingual corpus, dictionaries and rules.
As provided with source sentence it is processed through
many linguistic phases for collecting all the grammatical
information, and then words are disambiguated to gener-
ate target sentence. The syntax is mapped with the help
of parsing. It depends on colossal lexicons and linguistic
rules. The translation quality can be enhanced by adopting
Machine Translation Systems for Indian Languages: Review of Modelling Techniques, Challenges,…
contextual reference of a word in the sentence. This tech-
nique supersedes the MTS default settings of linguistic
rules of each word. The MTS using rule-based technique
can be build using the dictionary, transfer and interlingua
approach.
1. Dictionary-based MT This technique provides direct
transfer of meaning from source to target language
words based on dictionary entries. It does not cover
syntactic structure and semantic information of the lan-
guage pair. It is one of the simplest and easiest tech-
nique. It identifies the root word with the removal of suf-
fixes and lookup the bilingual dictionary for the meaning
of source word into the target word. The final output can
also be post-edited by re-ordering the sentence. [18]
2. Transfer-based MTS This technique transfers the source
language parse tree into target language parse tree. It
covers the syntactic aspect of the language pairs. It
performs the analyses of source language structure and
transfer it to the target language and finally generates the
target sentence [19].
3. Interlingua-based MTS This technique converts the
source language into an intermediate or interlingua
structure to generate target translation. It analyses the
source sentence and performs synthesis to convert it into
the target sentence. The idea is to capture the meaning
in the interlingua. It is independent of the language pair.
The transfer of interlingua is performed on the semantic
level as well as syntactic level [20].
Fig. 2  Rule-based modelling for different languages
RBMT is one of the oldest modelling technique applied to
develop MTS. Some of the recent research works using a
rule-based approach have been studied along with the evalu-
ation measures and are shown in the Table 2. The Table 3
exhibits methodology used by RBMT systems along with
their domains. The analytical study performed for MTS
build using rule-based approach concludes that most of
RBMT are for English to Hindi language and English to
Marathi as shown in Fig. 2.
2.2 Machine Engineered Translation
These modelling techniques involve more machine interven-
tion as compared to human-engineered as all the modules of
the system require machine processing. It builds a translation
model using machine learning from the corpus (monolin-
gual and bilingual) . The machine-engineered MTS include
corpus-based modelling technique which has been further
classified into statistical and neural-based.
2.2.1 Corpus‑Based MTS
The approach depends entirely on the corpus (bilingual,
multi-text or parallel corpus). MTS can be developed by
training the corpus to perform translation. In comparison to
the rule-based modelling technique, it requires fewer efforts
as it is more machine-dependent. This technique has been
further classified in statistical and neural-based.
13
 M. Singh et al.

Table 2  RBMT system based on their language pair with its respec- In the decoding phase, a source sentence chooses a
tive accuracy translation sentence with maximum probability as in
Author Year Language Parameters/accu- Eq. (3),
racy
argmaxP(t|s) = argmaxP(t) × P(s|t) (3)
BLEU Accuracy
The decoding phase substitute phrases from left to right
Kavirajan et al. [21] 2017 English–Tamil NA 71.80
to produce the translation. It provides a dynamic pro-
Rana and Antique 2016 English–Hindi NA 81.70
[22] gramming solution by applying the beam search algo-
Darbari et al. [23] 2015 English–Hindi 64.6 NA rithm. The fluency of translation output depends on the
Garje et al. [24] 2014 English–Marathi 44.29 49.78 language model and adequacy of translation is derived
Dubey [25] 2014 Hindi–Dogri NA 95 by translation model. Though decoding is a complex
Basavaraddi et al. 2014 English–Kannada NA NA process, it provides the output sentence along with re-
[26] ordering .
Adapanawar et al. 2013 English–Bengali NA 82.92   There are different models of SMT based on segmen-
[27] tation of source sentence into words [39, 40], phrase
Adapanawar et al. 2013 English–Marathi NA NA [41], syntax [42, 43] and hierarchy [44]. The SMT sys-
[27] tems better utilise data and human resources. It is bet-
Pisharoty et al. [28] 2012 English–Marathi NA NA ter than rule-based as it does not require human-formed
Nair [29] 2012 Malayalam–English NA NA linguistic rules which are time-consuming and language-
Batra and Lehal [30] 2010 Punjabi–English NA 85.33 specific. It demonstrates to have higher productivity and
Rajan et al. [31] 2009 English–Malayalam NA 79.60 quality for domain-specific translations such as news,
Sinha [32] 2009 English–Hindi 34.12 NA official documents and literature.
English–Urdu 35.44 NA   The survey performed on some of the recent research
Sinha and Jain [33] 2003 English–Hindi NA 90 work using statistical approach have been presented in
Dave et al. [34] 2001 English–Hindi NA NA the form of subsequent tables. The Table 4 compares
various SMT based MTS in terms of different evalua-
tion measures used for computation of accuracy. SMT
1. Statistical machine translation (SMT) modeling tech- systems are developed using different models having
niques. several toolkits. These are described in the Table 5. The
  It produces translation based on the statistical model corpora details along with their domains are mentioned
while learning from parallel and monolingual corpus. in the Table 6. The analytical study performed for indic
It was introduced in the late 1950’s [37] and had been languages using SMT modelling technique have been
performing quite well until the present time. The statisti- shown in Fig. 3 for various languages and concluded
cal model assumes the presence of aligned large quantity that maximum work has been done to translate different
and high-quality data. It encodes the data information languages into English and Hindi language.
in the language and translation model which is decoded 2. Neural machine translation (NMT) modelling NMT is
by the decoder to generate the translation. The model of the most recent technique for MT and is said to make
language and translation are produced using the given a substantially more accurate translation. It depends on
source and target sentence. The language model is built the model of neural systems in the human cerebrum and
using the monolingual corpus of the target language. It with data being sent to various “layers” to handle before
also assigns the probability to each string by calculating output. It utilises deep-learning procedures to guide
the relative frequencies. While the translation model is itself to translate content given existing reference trans-
built using parallel corpora by assigning probability, i.e. lation and build models. It forms the technique faster as
source sentence is a translation of the target sentence. it requires a single sequence model rather than multi-
Mathematically, using Bayes theorem in Eq. (1) [38], ple models as in SMT. It also produces a higher quality
P(s) × P(t|s) output. It models the source sentence s1 , s2 , s3, … , s(m)
P(s|t) = (1) to target sentence t1 , t2 , t3, … , tn with conditional prob-
P(t)
ability modelling context vector c as in Eq. (4)
The highest probability sentence is chosen as the best n
∑
translation using Eq. (2) logP(t|s) = logP(yi |y < i, c) (4)
i=1
emax ̂
= argmaxP(s)P(t|s) (2)

13
Machine Translation Systems for Indian Languages: Review of Modelling Techniques, Challenges,…

Table 3  RBMT system methodology adopted with its specific domains and corpus
Author name Year Methodology Domain Corpus

Kavirajan et al. [35] 2017 Sentence simplifier, linguistic rules NA 250 sentences
Rana and Antique [22] 2016 Fuzzy rule-based translation NA 50, 100, 200, 300, 500, 1000 sentences
Darbari et al. [23] 2015 TAG & MCSSG approach Rajya Sabha using Tree Adjoin- Rajya Sabha website since 2006
ing Grammar (TAG)
Basavaraddi et al. [26] 2014 Re-formatting, pre-editing, morpho- Goverment and education sector NA
logical analysis, transfer of internal
representation and generation along
with reformatting
Garje et al [24] 2014 Pre-translation processor,parsing, Tourism 1000 sentences from TDIL
named entity tagger, rearrangement
generator, sentence filter,word by
word translation,disambiguator,
target generator
Adak [36] 2014 Morphological analysis, lexicaliza- General 2574 words
tion: pos tagging, re-ordering,
transliteration and combine
Dubey [25] 2014 Direct technique NA 18,500 words
Pisharoty [28] 2012 Tokenization, lemmatization, pars- NA NA
ing, syntax validation, semantic
validation, translation, transforma-
tion and reconstruction of sentence
Nair [29] 2012 Pre-processing, morphological NA NA
parser, transfer and generator
Adapanawar et al. [27] 2013 Tokenisation, POS tagging, diction- Assertive sentences NA
ary lookup and rule extraction from
database
Batra and Lehal [30] 2010 Three components based approach: NA 500 sentences
analyzer, a transfer component, and
a generation component
Rajan et al. [31] 2009 Transfer link rules, morphological Word dictionary 5000
rules
Sinha [32] 2009 Interlingua based rule-based NA 100,000 words
approach.
Sinha and Jain [33] 2003 Rule based NA NA
Dave et al. [34] 2001 English analyzer with disambigua- Political and stock market stories 180 sentences (Ministry of Informa-
tion, UNL conversion and Hindi tion and technology) and Brown
generator corpus

The basic form of NMT consists of encoder and decoder
components. The encoder encodes the source sentence
into a context vector c while decoder decodes this
context vector and generate one word at a time. NMT
requires minimal domain knowledge. The sequence to
sequence learning proposed by [59, 60] employed by
[61, 62], i.e. reads a sentence till the end and output
one word at a time. It produced good translation results.
It has the drawback of encoding source sentence into
a fixed-size vector which deteriorated in quality when
exposed to longer sentences. However, this drawback
can be overcome by attention mechanism [63]. There
are two different architectures for constructing NMT.
1. Recurrent neural network (RNN) It has been pro-
ducing good quality translation result. RNN is com-
posed of encoder and decoder with similar working
of sequence to sequence learning. Different RNN
architecture are experimenting different models, [62,
64–69].
2. Convolution neural network (CNN) It has achieved
surpassing results for the word-based MTS, but
along with RNN [64, 70]. These work applied con-
volution layer on the bottom of the recurrent layer
which hinders the performance. The bottleneck
was handled by implementing the fully convolu-
tional model as suggested by [71, 72]. Later formed

13

Fig. 3  Statistical modelling for different Languages
stacked convolutional layers [73] with GPU version
of the neural model. The performance and accuracy
was improved with a number of models [72–75].
  NMT makes it easier to train large models and gen-
eralise long sentences. It doesn’t have to store phrase
tables, language models, score tables as in SMT.
  The survey performed on some of the recent research
work using neural-based approach has been presented
in the form of subsequent tables. There are differ-
ent toolkits available for developing the NMT system
experimenting with different methodology as in Table 7.
The experiment is conducted on different corpus and
domains mentioned in the Table 8. These systems are
evaluated with different evaluation measures such as
automatic or human, which are presented in the Table 9.
The analytical study performed for Indic languages
using neural-based techniques has been shown in Fig. 4
and concludes that English to Hindi translation systems
is more prominent as compared to other languages. This
is because more parallel corpora exist for this language
pair than other Indic languages.
2.3 Hybrid Machine Translation (HBMT) Modelling
HBMT is a preferred translation technique as it combines
the best of human-engineered and machine-engineered
approach. It is characterised by the use of multiple MT
modelling techniques within a single MT system. The
motivation for developing a hybrid approach stems from
the failure of any single technique to achieve a satisfactory
13
M. Singh et al.
level of accuracy. Many HBMT systems have been suc-
cessful in improving the accuracy of the translation
systems.
The HBMT architecture is guided by the human-engi-
neered approach, i.e. rule-based using corpora to build
[86–93], using corpus-based tool for weighing the RBMT
output [94–99], RBMT is guided by statistical post-edit-
ing [100–103]. Corpus-based HBMT uses rules at pre-
processing and post-processing suggested by [104–109],
incorporating dictionaries and rules in corpus-based MTS
[110–117] and building HBMT using corpus with statisti-
cal approach [118–124]. These several works demonstrate
that HBMT provides better translation quality. Most of
the techniques concatenate rules and data, whereas fewer
works are combining machine-engineered approaches.
Some of the recent work incorporates additional informa-
tion to guided RBMT or guided corpus-based technique.
The hybridisation technique has grown to speech trans-
lation, cross-lingual information retrieval and computer-
aided and post edited systems.
The survey performed on some of the recent research
work using the hybrid approach has been presented in the
form of subsequent tables. The HBMT developed for Indic
languages using different experiment or methodology along
with toolkits is depicted in Table 10. These system devel-
oped with corpus along with domains is mentioned in the
Table 11. Evaluating these systems with different evalua-
tion measures is shown in Table 12. The analytical study
performed for Indic languages using hybrid approach has
been shown in Fig. 5 and concludes that English to Marathi,
English to Hindi, Bengali to Hindi, and English to Punjabi
have notable more systems as compared to other languages.
Machine Translation Systems for Indian Languages: Review of Modelling Techniques, Challenges,…

Table 4  SMT systems for Author Year Language Parameters/accuracy

languages with its respective
accuracy BLEU NIST TER

Subalalitha et al. [45] 2018 English–Hindi NA NA 73.43

Jindal [46] 2018 English–Punjabi 87.67 NA NA
Patel et al. [47] 2018 English–Malayalam 8.25 NA 21.57
English–Hindi 19.43 NA 37.77
English–Punjabi 23.09 NA 44.06
English–Tamil 7.56 NA 23.62
Khan et al. [48] 2017 Bengali–English 19.7 3.786 NA
Hindi–English 19.3 37.79 NA
Malayalam–English 11.1 NA NA
Tamil–English 12.8 NA NA
Telugu–English 14.2 NA NA
Urdu–English 24.7 4.26 NA
Patel et al. [49] 2016 Bengali–Hindi 31.3 6.802 46.06
Marathi–Hindi 38.71 7.369 41.38
Telugu–Hindi 28.51 6.268 50.8
Tamil–Hindi 19.19 5.012 62.9
English–Hindi 20.75 5.616 61.7
Patel and Pimpale [50] 2016 Bengali–Hindi 33.77 7.195 45.52
Marathi–Hindi 41.2 7.935 39.25
Telugu–Hindi 29.72 6.669 49.26
Tamil–Hindi 20.38 5.34 62.25
English–Hindi 24 6.121 58.99
Das and Baruah [51] 2014 Assamese–English 11.32 NA NA
Ali et al. [52] 2014 English–Urdu 9.035 NA NA
Khan et al. [53] 2013 English–Urdu 2.9 6.36 NA
Ali et al. [54] 2013 English–Urdu 37.1 NA NA
Kumar and Kumar [55] 2013 Punjabi–English NA 97 NA
Ali et al. [52] 2010 English–Urdu 9.03 NA NA
Anwar et al. [56] 2009 Bangla–English NA 92.53 NA
Udupa and Faruquie [57] 2005 English–Hindi 13.445 4.5741 NA

3 Challenges of MTS for Processing Sanskrit
Language
This section discusses the types of challenges that one must
consider in developing an MTS. We examine these chal-
lenges along two dimensions, the first on different types
of linguistic considerations (e.g. syntactic word order and
semantic ambiguity) and the second on different types of
operational or technical considerations.
3.1 Linguistic Challenges
The pattern of divergence between two languages needs
to be recognised before building an MTS. The challenges
for translation is specific to language pair as it is dif-
ficult to build a general approach applicable to all lan-
guages. The work studying the deviation of languages by
Dorr [133, 134] has formed a basis for further research
for Indian languages. The challenges to translate Eng-
lish–Sanskrit–Hindi MT [135], English–Hindi [34] and
English–Sanskrit [136] are also considered. These linguis-
tic challenges are specific to Sanskrit language and should
be considered before building MTS for Sanskrit to English
language or Sanskrit to Hindi language.
• Sanskrit contains complex or compound words being
influenced by oral tradition i.e continuous strings of
characters without word boundaries or punctuations.
It becomes difficult to guess the boundaries as they
undergo euphonic changes [9].
• In Sanskrit, there is a special category of verbs requir-
ing special treatment termed as thematic divergence.
The subject Noun Phrase (NP) in Sanskrit and Hindi is
the dative case while in English is the nominative case
which causes a divergence in translation [135].

13
 M. Singh et al.

Table 5  SMT models along with methodology and toolkit

SMT model Author Year Methodology Toolkit

Phrase-based Subalalitha [45] 2018 n-gram and Naive Bayes probability NA

Shishpal Jindal [46] 2018 IRSLTM language model, trained model with Moses, IRSLTM and GIZA++
GIZA++ alignment and testing
Patel et al. [47] 2018 Pre-processing, re-ordering, suffix seperation, Moses, MERT, KenLM and Kneser-Key
transliteration
Khan et al. [48] 2017 Sampling, tokenization, tuning set Moses, Giza++, Kneser-Ney, SRILM
Patel and Pimpale [49] 2016 Pre-processing, transliteration Modified KenLM, Moses
Patel et al. [50] 2016 Suffix separation (compound splitting and Modified KenLM
reordering)
Ali et al. [52] 2014 Manual alignment, translation and tuning Moses , Giza++
Das and Baruah [51] 2014 Text, decoder and transliteration IRSTLM tool, Giza++, Moses Decoder
Ali et al. [54] 2013 Tokenization, training and tuning Moses, IRSTLM
Kumar and Kumar [55] 2013 N-gram model and transliteration NA
Ali [52] 2010 Data distribution, mkcls, GIZA++ and Moses, GIZA++ and MERT
MERT tuning
Anwar et al. [56] 2009 Tokenization, syntax analysis, parsing and NA
NLP conversion
Anwar et al. [56] 2009 Tokenization, syntax analysis, parsing and NA
NLP conversion
Cherry [58] 2008 NA Moses, Giza++, and phrasal
Udupa and Faruquie [57] 2005 NA Moses, Giza++, and Phrasal
Udupa and Faruquie [57] 2005 NA IBM models 1, 2, and 3
Hierarchical-based Khan et al. [48] 2013 Sampling, tokenization and tuning Moses Decoder, Giza++, SRILM

Table 6  SMT classification based on domains and corpus

Author name Year Domain Corpus

Subalalitha [45] 2018 News, agriculture and technical phrases IIT Bombay and manually collected corpus
Jindal et al. [46] 2018 Health, tourism and Gyan nidhi Manually curated
Patel [47] 2018 General MTIL-2017
Khan et al. [48] 2017 Consumer, education, health, housing, legal and EMILLE
social documents
Patel et al. [49] 2016 Health, tourism, and general domain ILSMT and 23K sentences from other
Patel and Pimpale [50] 2016 Health, tourism, and general domain ILSMT, 23K, and 500 sentences from other
Das and Baruah [51] 2014 Tourism data Parallel corpora of about 8000 sentences
Ali et al. [52] 2014 Quran, ahadeeth 6000 sentences
Ali [52] 2010 Religious Adaheeth
Khan et al. [48] 2013 NA EMILLE
Ali et al. [54] 2013 NA 41,208 sentences
Kumar and Kumar [55] 2013 Names 15,000
Anwar et al. [56] 2009 NA NA
Udupa and Faruquie [57] 2005 News, government documents, conversation, 150,000 sentence pairs
and magazine articles

• Sanskrit is rich in inflectional and morphology. This • Almost all Indian Languages vocabulary has been
richness makes a difference in the last character of the derived from Sanskrit. There have been cases of mean-
word, its gender and makes difficult to remember dif- ing shift, reduction and expansion. It makes difficult to
ferent forms of word inflections [9]. understand the Sanskrit text without the prior knowledge

13
Machine Translation Systems for Indian Languages: Review of Modelling Techniques, Challenges,…

Fig. 4  Neural modelling for different languages

Table 7  NMT system based on toolkit with its respective methodology

Toolkit Author name Year Methodology Model

Tensorflow Jha et al. [76] 2018 Sequence to sequence, alignment, hierarchical Bi-LSTM with attention
attention network, transformer network and
character encoding
GoogleTranslate API Choudhary et al. [77] 2018 Sequence to sequence model, attention, BPE, Bi-LSTM with attention
word embedding
OpenNMT Pathak and Pakray [78] 2018 Data pre-processing, training, encoding, NA
decoding and translation
OpenNMT Ramesh and Sankaranayanan [79] 2018 NA Attention-based model
TensorFlow Singh et al. [80] 2018 Neural machine translation, training, testing LSTM with attention
Nematus Jigar Mistry [81] 2017 Encoder–decoder model, attention and BPE Vanilla LSTM or GRU​
OpenNMT Revanuru et al. [82] 2017 Input, Bi-LSTM, Sum, LSTM, bridge and Deep Bi-LSTM
decoder
OpenNMT Tennage et al. [83] 2017 Pre-processing, OpenNMT, benchmark, train- NA
ing and word phrases
TensorFlow Aggarwal and Sharma [84] 2017 LSTM and GRU cells, encoders and decoders Bi-LSTM with attention mode
Theano Yerra et al. [85] 2016 NA Attention-based model

Table 8  NMT systems with their respective domain and corpus

Author Year Domain Corpus

Choudhary et al. [77] 2018 News, Bible, Cinema, Movie Subtitles EnTamV2.0 and Opus
Jha et al. [76] 2018 Dictionary Manually curated words from dic.
Pathak and Pakray [78] 2018 General domain MTIL
Ramesh and Sankaranayanan [79] 2018 Wikipedia Wikimedia dumps
Singh et al. [80] 2018 Health, tourism, agriculture and entertainment TDIL, EMILLE, OPUS
Revanuru et al. [82] 2017 Agriculture, entertainment, health and tourism TDIL-DC
Aggarwal and Sharma [84] 2017 General domain 50,000 sentences from Bojar corpus and ILCI
Tennage et al. [83] 2017 Annual reports, establishment codes, order Official government documents of Sri Lanka
papers, and official letters
Jigar et al. [83] 2017 Health and tourism ILCI

13
 M. Singh et al.

Table 9  NMT system with their Author name Year Language Parameters/accuracy

language pairs and respective
accuracy BLEU Accuracy

Jha et al. [76] 2018 Hindi–Bhojpuri 90.89 90.23

Choudhary [77] 2018 English–Tamil 8.33 NA
Pathak and Pakray [78] 2018 English–Hindi 52.54 NA
Ramesh and Sankaranayanan [79] 2018 English–Tamil 5.53 NA
English–Hindi 3.97 NA
Singh et al. [80] 2018 English–Punjabi 26.07 NA
Mistry [81] 2018 English–Hindi 26.88 NA
Bengali–Hindi 33.87 NA
Gujrati–Hindi 53.95 NA
Revanuru et al. [82] 2017 Punjabi–Hindi 46.47 NA
Gujrati–Hindi 35.69 NA
Urdu–Hindi 22.47 NA
Tamil– Hindi 7.56 NA
Aggarwal and Sharma [84] 2017 English–Hindi 9.23 NA
Tennage et al. [83] 2017 Sinhala–Tamil 7.5 NA
Tamil–Sinhala 12.75 NA
Yerra et al. [85] 2016 Bengali–Hindi 20.41 NA

Fig. 5  Hybrid modelling for different languages

of original meaning. There are various trends in Sanskrit
literature for commentaries. The presentation of com-
mentaries in the Sanskrit language is in a nested form
which makes it difficult to understand for modern schol-
ars [9].
• The structure difference between languages lead to
problems in translation. Sanskrit and Hindi are Vib-
hakti and karaka based languages while in English
whenever Noun Phrase is encountered Vibhakti place
is replaced with prepositions or null [135].
• Sanskrit consists of complex words resulting in the
translation of two to three words of English in the one-
word translation of Sanskrit. This Conflational and
Inflation divergence is encountered for Sanskrit and
Hindi translation as well [135].

13
Machine Translation Systems for Indian Languages: Review of Modelling Techniques, Challenges,…

Table 10  Hybrid MTS modelling technique based on its model and toolkit

Technique Author Year Model Toolkit

Rule-based + statistical-based Salunkhe et al. [126] 2016 NA Open NLP

Rule based + statistical phrase based Dhore and Dixit [129] 2011 NA NA
Rule-based (lattice based lexical transfer) + statistical-based Chatterji et al. [130] 2011 Phrase based NA
Rule-based (lexical transfer based) + statistical-based Chatterji et al. [132] 2009 Phrase based Giza++
Statistical machine translation + translation memory Nithya and Joseph [127] 2013 Phrase based IRSTLM,
GIZA++,
Moses
decoder
Rule-based + example-based Kaur and Laxmi [128] 2013 NA NA
Rules-based + neural-based) Shahnawaz and Mishra [131] 2011 NA java(jdk1.5)
with Mat-
lab 7.1
Rule-based + example-based + statistical-based Dhariya et al. [125] 2017 Phrase based NA

Table 11  Hybrid systems with Author Year Domain Corpus

its domain and corpus
Dhariya et al. [125] 2017 NA CFILT, IIT-Bombay
Salunkhe et al. [126] 2016 NA Parallel Corpus (IIT-Bombay)
Nithya and Joseph [127] 2013 Indian history and 563 sentences
Islamic history
Kaur and Laxmi [128] 2013 News headlines 300 sentences
Dhore and Dixit [129] 2011 Banking glossary Reserve Bank of India
Chatterji et al. [130] 2011 Tourism 2000 sentences
Shahnawaz and Mishra [131] 2011 NA NA
Chatterji et al. [132] 2009 Written EMMILE-CIIL

Table 12  Hybrid MT based on Author Year Language Parameters/accuracy

language with its respective
accuracy BLEU Accuracy

Dhariya et al. [125] 2017 Hindi–English NA 86.50

Salunkhe et al. [126] 2016 English–Marathi NA 83
Nithya and Joseph [127] 2013 English–Malayalam 69.33 75.30
Kaur and Laxmi [128] 2013 English–Punjabi NA 81.67
Dhore and Dixit [129] 2011 English–Hindi NA 97.25
English–Marathi NA 97
English–Gujarati NA 96.50
Chatterji et al. [130] 2011 Bengali–Hindi 29.45 NA
Shahnawaz and Mishra [131] 2011 English–Urdu 69.54 NA
Chatterji et al. [132] 2009 Bengali–Hindi 22.57 NA

• Though Sanskrit is a free-word order language still, there
are cases where the change in the word order changes the
meaning of a sentence for translation of language pair
Sanskrit and English [135].
• Sanskrit and Hindi language pair have most commonly
passive voice while English has active voice sentences.
This change of voice leads to problems in translation for
specific language pair [135].
• There is categorical and lexical divergence for Sanskrit–
English and Sanskrit–Hindi language pair. The categori-
cal divergence occurs in case of mismatch in parts of
speech of translation pair languages. In the translation
process when an exact match is not mapped from one
language to another leads to lexical divergence. In San-
skrit different meaning is generated with the addition of
upsarga to the verb [135].

13

• In the translation of Sanskrit to English/Hindi there are
different adjuncts and clauses are located causing diver-
gence in translation. This divergence changes the sen-
tence construction of language pair [137].
• Sanskrit has honorific features containing verb endings
with adjectives and noun. These plural verb and plural
pronouns inflections are caused because of socio-cultural
aspects of languages [137].
• The mapping of time from English to Sanskrit and Hindi
causes a problem as am and pm in English cannot be
mapped in Sanskrit as afternoon and morning. The term
am and pm cannot be translated as such in Sanskrit and
Hindi [137].
• There is formal grammar defined by Panini for Sanskrit
but no such parallel grammar exist for Hindi or other
European languages. In the absence of such parallel
grammar, exceptional cases are covered by forming lin-
guistic rules. There are cases were Vibhakti diverges
from Sanskrit to Hindi such as optional, exceptional, dif-
ferential, alternative, non-Karaka, verbal and complex-
predicate divergence [137].
3.2 Technical Challenges
This section discusses the technical challenges of Sanskrit-
based MTS. The challenges are encountered while develop-
ing and applying modelling techniques for MTS.
• Domain mismatch In NMT based system if the content is
not from the same domain, then it exhibits poor perfor-
mance. It has low-quality for out of domain text, as for
fluency it sacrifices adequacy. The translation misguides
the user by visualising the fluent output in case of infor-
mation gisting [138, 139].
• Amount of training data It performs well for high-
resource languages as compared to low-resources lan-
guages as its learning depends on the amount of training
data. NMT systems trains million of data, showing direct
proportionality to accuracy [140, 141].
• Noisy data NMT system is not robust to noisy data in
corpora such as misaligned sentences [142], poorly trans-
lated sentences, content in wrong languages. In such
cases, NMT fails to predict the relationship between the
language model and the input data context. Even vary-
ing the training ratios, the problem does not dissolve as
it produces inadequate output.
• Word alignment Aligning input words to output words
was served by the attention model of NMT [63]. The
performance of attention based NMTs is very poor in
case of more substantial sentences, and it does not pro-
vide accurate word alignment. Incorporating discrete
translation and lexicon dictionaries [143] for improving
system with fertility and coverage modelling [144]. The
13
M. Singh et al.
attention mechanism may produce better alignment if
provided with guided learning [145, 146]. The source
sentence can be translated into many target sentences in
SMT as well. Aligning of phrases in the target concern-
ing source becomes cumbersome. An efficient alignment
algorithm is required after translation. There are several
techniques such as a template for aligment [118, 147],
Hidden Markov Model (HMM) [148], toolkits [149–
151]. Despite various attempts to enhance alignment in
SMT systems, it does not fulfil the role and shows diver-
gence.
• Beam search The decoding process generates the trans-
lation of highest probability. To find the highest prob-
ability translation, a search operation is performed from
all possible translations. In NMT, the size of all possible
translation is termed as beam size for each input word. It
is not directly proportional to the accuracy of the system
as after a point it starts degrading. It requires manually
normalizing of scores by the sentence length. Falling out
of the optimal range (30–50) of beam size the translation
quality starts degrading. The wider beams deteriorate the
quality with shorter translations [152].
• Longer sentences and inflectional category words: Com-
plex and large sentences quality is low as compared to
small sentences. Even words which are of low frequency
are not easily translated with NMT based systems [152].
• Parallel Corpus Corpus-based modelling techniques
require a large amount of parallel and monolingual cor-
pus but it is costly and time-consuming. Training data is
directly proportional to the model constructed. A collec-
tion of millions of monolingual sentence yields a better
language model producing fluent output. The translation
model is trained using parallel corpus producing the ade-
quate translation. Even with huge corpora, translation
quality is very coarse. This modelling technique does not
apply to many Indic languages as there is very less or no
parallel corpora for some low-resource languages [152].
• Idioms and multiword expressions The properties exhib-
ited by idioms and multiword expression make it difficult
for them to translate. The corpora should be customised
for a specific idiom for particular language pair to be
most effective. Even then, SMT requires pre-processing
and post-processing steps for handling such cases [153].
• Time consuming SMT, RBMT can also be expensive as it
requires a lot of upfront costs. In this, both the processes
i.e, pre-processing and corpus creation are expensive
and time-consuming. It also requires collaboration with
computer scientists, translators, linguists and statisticians
[154].
• Learning For the learning of system every phase requires
continuous error detection. It is harder to fix mistakes
in the system once they have been implemented. With
models like RBMT, you can fix errors and remove cer-
Machine Translation Systems for Indian Languages: Review of Modelling Techniques, Challenges,…
tain words quite easily. With SMT, you need to retrain
the whole system and check if other errors have been
emerged or not [153].
• Linguistic issues Even training the SMT system with 100
million words produce a partially excellent translation.
It suffers from various linguistic issues such as man-
gled grammar, wrong word choices, name translation,
unknown words and syntactic transformations [155]
• Linguistic knowledge Some linguistic information still
needs to be set manually (such as rules, part of speech).
This requires human intervention and linguistic knowl-
edge of the source and target language [155].
• Generalized system It is hard to deal with rule interac-
tions in big systems, resolve ambiguity and handle idi-
omatic expressions. It is difficult to build a generalized
system handling all the aspects of a language pair [154].
• Domain adaptability Although RBMT systems usually
provide a mechanism to create new rules and extend and
adapt the lexicon. Adapting to new domains requires
extensive knowledge of language and human-effort as
each word requires a rule to disambiguate its meaning
[154].
• Resources The resources required for developing RBMT
are linguistic rules, dictionaries, language-specific tools,
morph-analyser, parser and generator. It is an expensive
process as it requires much human effort and knowledge.
There is also a requirement of efficient corpora for a par-
ticular language. Some languages do not have sufficient
data and preparation of these data is a challenging task
and time-consuming process [154].
• Modification To change in existing RBMT system
requires to process as an alteration in the rules. It is more
expensive as compared to generating new rules [154].
4 Comparison of Various Modeling
Techniques
In MT research for Indic languages, RBMT and SMT are
the methods which are used most frequently. Though RBMT
is the oldest approach it can achieve good results but the
development is very time-consuming as manually linguis-
tic rules need to be fixed for every word in a sentence. In
terms of investment, the customisation cycle needs to reach
the quality threshold which can be quite long and costly
in terms of human resources. The RBMT systems are built
with fewer data as compared to SMT systems, along with
dictionaries and language rules to translate. It does not pro-
duce fluently translated output. Also, language is constantly
changing, which means rules must be managed and updated
wherever necessary in RBMT systems. Moreover, SMT sys-
tems require much less time and linguistic knowledge. SMT
models require more computer processing power and storage
capacity to build and manage large translation models. Fig. 6
shows that SMT systems are mostly used by the researchers
for the translation of different languages. On the other hand,
neural and hybrid-based systems are very less used for Indic
language pair due to insufficient parallel data. The perfor-
mance of these modelling techniques is good with fluent
output. The computational requirement for these techniques
is more than human resource requirements. Machine learn-
ing and linguistic knowledge apply to these techniques to
improve performance. These techniques sometimes exhibit
out of domain quality. The error analysis is difficult to per-
form for these techniques. So, these systems need to test
for future translations. The performance of some of the MT
modelling techniques for different quality parameters are
compared in Table 13. The year-wise usage of these tech-
niques is shown in Fig. 6.
5 Comparison of Machine Translation
Systems
MT aims to translate one language to another by utilis-
ing different resources over the years; different modelling
techniques have evolved to provide efficient translations.
(Fig. 7). The principal objective is to fulfil the language
gap between two distinct languages involving individuals,
groups or nations. In India, we have different languages and
web-contents extensions which needs massive language
translations. The Table 14 and 15 contains the classification
of MTS based on human and machine-engineered approach
along with their features and outcomes. This table contains
information like the modelling technique used, features of
the technology and outcomes along with different language
SMT
41%
22%
RBMT
23%
14%
Neural
Hybrid
Fig. 6  Overall percentage of various modelling techniques
13
 M. Singh et al.

Table 13  Comparison of Features RBMTS SMTS EBMTS NMTS

various MTS technique based
on various parameters Performance Good Medium Good Good
Fluency Less Medium High Medium
Robust Yes No Yes No
Human resources requirement High Less Medium Less
Computational resources requirement Less Medium Medium High
Machine learning applicability No Yes Yes Yes
Linguistic knowledge requirement Yes No Yes Yes
Use of grammar Yes No No No
Out of domain quality Medium Low High Low
Predic quality Good Similar Very well Similar
Consistency High Low Medium Low
GPU requirement No No No Yes
Language dependency Yes No No No
Maintenance Difficult Easy Easy Easy
Model size Huge Huge Moderate Small
Error analysis Easy Difficult Difficult Impossible
Parallel corpus requirement No Yes Yes Yes
Dictionary and rule requirement Yes No No No
Extendable Difficult Easy Easy Easy

Fig. 7  Modelling techniques based on the year of its development

13
 Table 14  Human-engineered systems developed for various languages
Author Corpus Language pair Features Outcome

Technique: Rule-based
Adapanawar et al. [27] General database English–Marathi Open source NLP tools are used for developing the system.
Database of rules using a bilingual dictionary is used for
mapping
Adak [36] Parallel corpus English–Bengali Use a soft computational technique where the fuzzy If-Then
rule is applied to choose a lemma from prior knowledge
Pisharoty et al. [28] General database English–Marathi For improving the performance of the system, grammar and
spell checker can be used. The sentiment analysis module
can also be used
Garje et al. [24] 1000 sentences English–Marathi Semantic and morphological properties are maintained in the
lexicon grammatical structure of the target language gives
importance for better translation
Basavaraddi et al. [26] General database English–Kannada Complex morphology of the target language handled by the
morphological generator. Syntax reordering overcomes
The research represents a theoretical and grammatical frame-
work which is extendable
The proposed system works well in sentence translation from
English to Bengali and they obtain 82.92% F-measure on the
basis of their test case analysis
The additional functionalities have improved the system accu-
racy, although there is a trade-off of time
The accuracy of the system using TDIL corpora is 44.29% and
for human translation is 49.78%
Differences were found in the syntactic module (word order and
morphological level)
Machine Translation Systems for Indian Languages: Review of Modelling Techniques, Challenges,…

syntactic differences
Technique: Transfer based
Nair and Peter [29] 1000 sentences Malayalam–English Artificial intelligence techniques used for system develop- The system was tested for 1000 different sentences and reported
ment. Splitter for splitting compound words, bilingual true result for the sentences which had two subordinate
dictionary, the morphological parser clauses. The system is easily extendable for other language
pairs

13


13
Table 15  Machine-engineered systems developed for various languages
Author Corpus Language pair Features Outcome

Technique: Statistical based

Kumar and Kumar [55]
Ali et al.
Technique: Phrase-based statistical
Pingali and Vasudeva [156]
Technique: Hierarchical phrase-based
Khan et al. [53]
Technique: Hybrid
Nithya and Joseph [127]
Dhore and Dixit [157]
Nair et al. [29]
Godase and Govilkar [158]
Singla and Baghla [159]
Technique: Example based
Sinhal and Gupta [160]
Anuj and Manoj Kumar [161] General database
1000 names Punjabi–Engish System has training to learn and transliteration The accuracy achieved by the system is 97% test set gives
the BLEU score 32.11
20,173 sentence pair English-Urdu Moses is used for language training with modeling NA
toolkit IRSILM
43,500 sentences Telugu–English Modules: language model, translation model, and The system exhibits based on geometric average fluency
decoder of 2.693 and adequacy of 2.93
EMILLE Corpus English–Urdu EMILLE corpus is used. On Urdu monolingual corpus BLEU score for the system in fivefold test data, 40%
(12,500 sentences) language model is built. Using SRILM toolkit N-gram (phrase-based) and 29% (hierarchical-based) NIST
model is used score for the system is 73% (phrase-based) and 63%
(hierarchical-based)
563 sentences Malayalam–English A statistical method is applied to the corpus and apply- BLEU score for the baseline system was 68.14 and for the
ing machine learning techniques for translation hybrid system was 69.33
1000 words English–Devnagari The multilingual dictionary is created using 1000 bank- The multilingual dictionary is improved and semantic
ing glossary, which is available on the RBI website. C rules in the parser design have also improved
language is used for the lexical analyzer. For running
the system Bison tool is used
General database English–Hindi New rules has been added The proposed system design shows the accurate results
than other systems that are 94%.
Dictionary database English–Marathi Bilingual corpus is used for training. Parsing is used by It makes resources available to everyone by presenting
the system a complete architecture and several algorithms for the
system
15,000 sentences English–Punjabi Using rule-based modelling technique parsing is per- The accuracy achieved by the system is 81.67%
formed by the system
677 sentences English–Hindi Comparing sentence to extract the translation. Parallel The system provides 96.07% word strength and 86%
corpus is used for Training. Uses various modules such precision for 677 sentences
as similarity matrix, training matrix, tagging matrix
Malayalam–English Phases of the system are an acquisition, matching, Best translation quality is given by the 75% test and reor-
and recombination searching mechanism is used for dering problems
searching fragments of Malayalam
M. Singh et al.
Machine Translation Systems for Indian Languages: Review of Modelling Techniques, Challenges,…

pairs. An analysis of the percentage of work performed for many languages. Sanskrit is the mother tongue of 24,821
different languages is shown in Table 16. In the field of people and Hindi of 52,83,47,193 people, i.e., 43% of total
MT, more systems have been developed to translate Hindi languages in India according to the census of India [162].
and English language as per analysis shown in Fig. 8. The Sanskrit is considered as the donor of almost all Indian
paper contains details of research work for other ancient languages [163]. The vast reserves in the Sanskrit lan-
Indian languages such as Punjabi, Bengali, Marathi, Tel- guage can be converted into other languages [164].
ugu, Tamil, Assamese, Urdu, Malayalam, Gujarati, San- The rich knowledge base in Sanskrit is its grammatical
skrit, Kannada, Dogri, Sinhala and Devnagari in Tables 14, tradition attracting Indian and western scholars. It is one
15, 16 and concludes that work for Sanskrit language is of the spoken languages and also at one time was known as
minimal despite of huge corpora and literature. ’Lingua Franca’ of the world intellectuals [165]. Sanskrit
has the text of different domains ranging from Ayurveda,
Philosophy and Astronomy. It holds a rich grammar con-
6 Machine Translation System for Sanskrit fined by Panini nearly 2500 years ago formulating 3949
and Hindi Language Processing rules, which extended later on [2]. Sanskrit has the strong-
est and simple non-ambiguous grammar [166]. Sanskrit
In a large multilingual society like India, there is a great has the richness of scientific literature with extensiveness
interest in the translation of text from one language to and comprehensive analysis, structured approach and tra-
ditional grammar [167]. Many people have attempted to
Table 16  Comparison of modelling techniques based on its develop- write a grammar for Sanskrit language using the Paninian
ment for Indic language pairs framework and used it to develop translation system [168].
Language pair SMT RBMT NMT Hybrid The Sanskrit grammar is termed as ‘Father of Informat-
ics’ as it builds a relationship between speech and utter-
Bengali–Hindi ✓ ✗ ✓ ✓
ance of speaker and meaning derived by the listener [169].
Marathi–Hindi ✓ ✗ ✗ ✗
Hence, the primary objective of Paninian grammar is to
Telugu–Hindi ✓ ✗ ✗ ✗
form a theory of human natural language communication.
Tamil–Hindi ✓ ✗ ✓ ✗
Sanskrit and Hindi belong to the same Indo-Aryan family
English–Hindi ✓ ✓ ✓ ✓
[137]. They both have structural and lexical similarity as
Assamese–English ✓ ✗ ✗ ✗
Hindi inherits from Sanskrit. Sanskrit has the rich and
English–Urdu ✓ ✓ ✗ ✓
structured grammar in the form of Panini Astadhayayi
Punjabi–English ✓ ✓ ✗ ✗
whereas in Hindi such parallel grammar does not exist.
Bengali–English ✓ ✗ ✗ ✗
Therefore, it becomes difficult to map the divergence
Hindi–English ✓ ✗ ✗ ✓
between these two languages. The non-existence of such
Malayalam–English ✓ ✗ ✗ ✓
grammar leads to exceptional cases which uncover lin-
Tamil–English ✓ ✗ ✗ ✗
guistic generalisations such as Vibhakti in Hindi. Despite
Telugu–English ✓ ✗ ✗ ✗
rich grammar, choosing Sanskrit as a source language is
Urdu–English ✓ ✗ ✗ ✗
difficult because parsing fails due to its synthetic nature
Bangla–English ✓ ✗ ✗ ✗
in which single word can run up to 32 pages. With the
English–Malayalam ✗ ✓ ✗ ✓
rich diversity of grammar, text and resources, it is per-
English–Tamil ✗ ✓ ✓ ✗
plexing to find access to Sanskrit computational tools.
English–Dogri ✗ ✓ ✗ ✗
One of the many reasons stated is unable to access the
English–Marathi ✗ ✓ ✗ ✗
literature as Sanskrit scholars do not turn towards com-
English–Bengali ✗ ✓ ✗ ✗
puter science. There are few systems for processing San-
English–Kannada ✗ ✓ ✗ ✗
skrit language and translating it to English or vice-versa.
Punjabi–Hindi ✗ ✗ ✓ ✗
These are depicted in Table 17. There are decidedly fewer
Gujrati–Hindi ✗ ✗ ✓ ✗
systems for translation of Sanskrit to Hindi as compared
Urdu–Hindi ✗ ✗ ✓ ✗
to English–Sanskrit as displayed in the Fig. 10. Differ-
Sinhala–Tamil ✗ ✗ ✓ ✗
ent modelling techniques have been used for processing
Tamil–Sinhala ✗ ✗ ✓ ✗
the Sanskrit language as depicted in the Fig. 9. Figure 9
English–Punjabi ✗ ✗ ✓ ✗
concludes that there is a lot of work done for RBMT sys-
English–Gujrati ✗ ✗ ✗ ✗
tems for the Sanskrit language, whereas no work is done in
English–Devnagari ✗ ✗ ✗ ✗
Sanskrit Translation using Statistical modelling technique.
English–Sanskrit ✗ ✓ ✗ ✓
The other techniques like example-based and hybrid-based
Sanskrit–Hindi ✓ ✗ ✗ ✗
are equally used for Sanskrit translation. The performance

13
 M. Singh et al.

33%

30 %

24%
Percentage(%)

20 %

10 %
7%
6%
5%
4% 4%
3% 3%
2% 2%
1% 1% 1% 1% 1%
0%
Telgu

Tamil

Gujrati
English

Hindi

Kannada

Sinhala

Dogri

Bangali

Bhojpuri

Assamese

Sanskrit

Punjabi

Malyalam

Marathi

Urdu
Languages

Fig. 8  MTS developed for various languages

Fig. 9  MTS developed for dif-

ferent modelling techniques Neural-based Approach 0%

Hybrid Approach 14%

Example-based Approach 14%

Statistical-based Approach 28%

Rule-based Approach 43%

0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50%

Percentage of modelling techniques for Sanskrit languages

Sanskrit-Hindi 37.5%

English-Sanskrit 62.5%

0% 10% 20% 30% 40% 50% 60% 70%

Percentage of Systems

Fig. 10  MTS developed for Sanskrit language

13
 Table 17  Comparison of different MTS for Sanskrit and Hindi language processing
Author Technique Language pair Corpus Features
Pandey and Jha [170] Statistical-based Sanskrit–Hindi 24,000 (bilingual) The system is being trained
25,000 (monolin- simultaneously on Microsoft
gual) Translator Hub (MTHub)
and is intended only for sim-
ple Sanskrit prose texts
Rathod et al. [171] Rule and example-based English–Sanskrit NA The text input is processed
with spell checker followed
by token generator, transla-
tor, parser, EBMT/RBMT
database and generator
Shukla and Shukla [172] Rule-based English–Sanskrit NA It unifies the isolated word
class under the speech
recognition type, tradi-
tional dictionary rule-
based machine translation
technique and text to speech
synthesizer
Bahadur et al. [173] Rule-based English–Sanskrit 500 sentences The sentences can be simple
and compound with the
affirmative and imperative
type or of active or passive
voice having any of the three this modeling technique for
tenses i.e. present, past, and extra large sentences
future
Jayan et al. [174] Example-based English–Sanskrit 125 input-output pairs Proposed a novel method
that uses rules and ANN
technique to detect and
implement the adaptation
rules for the divergence in
English to Sanskrit machine
translation
13
Results Issues
39.17 % for long, complex and The system takes input only in
compound sentences. 41.17 Devanagari Unicode script
for bilingual and monolin- and gives output in same.
gual sentences Sometimes the system does
not a response to long and
compound sentences. One can
translate Sanskrit sentence by
giving direct input in the box
An average improvement of EBMT is better than RBMT but
10% has been achieved by performance degraded in case
using EBMT in the transla- of extra large sentences
tion quality than RBMT
An average improvement of Translates only simple sen-
7% has been achieved in the tences, not complex sentences
translation quality
An average improvement of The sentence is correct in terms
13% has been achieved in of grammar, but the transla-
the translation quality. And tion is not correct. Few words
90% accuracy is achieved by in English may be used as
both noun and verb. This
generates ambiguity for the
system
Machine Translation Systems for Indian Languages: Review of Modelling Techniques, Challenges,…
An average improvement of For better improvement, future
8% has been achieved in the work is carrying to perform
translation quality case-based reasoning in a
combination with rule-based
and ANN model for this
purpose
 M. Singh et al.

of these techniques for the Sanskrit language is compared

out of the scope. Some among

discourse considerations were

aspect that has not been dealt

use of right prepositions and

with is the structuring of the

not handled, pragmatics and
in Fig. 10 and accordingly hybrid performs much better

them are the ambiguities of

plurals of words. Another

Semantic ambiguities were than other techniques by achieving 37% improvement as
compared to others.

sentence
7 Open Issues
Issues

MTS still face many issues. In this section, current as well

7% has been achieved in the

as future challenges are deliberated due to the extreme need

An average improvement of

for efficient MTS.

translation quality

• MT is an NP-hard problem designed to provide accurate

translations. The technology has improved drastically in
the past 10 years, but it is still a work in development.
Results

Therefore, even after post-processing, the meaning of the

original document is not 100% accurate.
• Developing a large amount of parallel data or corpus is
individual words of Sanskrit

mon occurrence in Sanskrit

required for application of statistical and neural tech-

the most important part of
into longer ones is a com-

texts, sandhi vichcheda is

niques for processing the Sanskrit language. Machine-

Since the conjunction of

engineered techniques require an enormous amount of

parallel text, even that is not available for specific lan-
the translator

guage pairs as some languages are not rich in lexical

resources [176].
Features

• For translating between the language pairs which have

different word order such as Subject Verb Object (SVO)
(such as in English) and Subject Object Verb (SOV)
(such as in Hindi), difficulties arise as output translation
does not follow the word order of target language. It is
also termed as structural divergences as the structure of
the two languages differs from each other. For example,
Corpus

He saw a man on the hill with a telescope. Here whether

English–Sanskrit NA

the man is seen with a telescope or the man is seen on the

hill is ambiguous. The post-editing rules can be used for
Language pair

handling such problems. However, ambiguity remains a

challenging problem in MTS [177].
• Translating one language into another also entails struc-
tural and semantic divergence. Handling of ambiguity
whether partial or complete at various levels such as lexi-
cal, semantic and structural need to be modelled.
• Handling various grammatical issues related to a par-
ticular language such as Subject-Verb, Adjective-Noun
Subramanian et al. [175] Rule-based
Technique

arrangement and parts of speech assignments.

• In RBMT approach, anaphora resolution problem arises,
which is handled by the expression whose contextual
interpretations depend on another expression. It also
includes problems such as zero-pronoun anaphora and
Table 17  (continued)

pro-drop anaphora [18].

• The case marking is the grammatical category (Noun,
pronoun, Adjectives) to a particular word or group of
Author

words which denote the course of action of that word

in sentence or phrase. The case marking is language-

13
Machine Translation Systems for Indian Languages: Review of Modelling Techniques, Challenges,…
dependent, and therefore it is different for different lan-
guages such as Sanskrit and Hindi.
• Handling of idiomatic expressions, rare words and out of
vocabulary words in which the meaning is not inferred
from the word are difficult to handle.
• In setting linguistic features of languages there are many
challenges such as a parser, root, GNP (Gender, Number,
Person), TAM (Tense–Aspect–Modality) and reordering.
8 Future Research Directions
From the literature and analysis of current MTS following
research directions are suggested.
• MTS helps to unite the world socially, culturally and
technologically. Hence, there is a significant require-
ment for inter-language translation for transfer and shar-
ing of information and ideas. Sanskrit is considered as
an essential language in the Indo-European families.
Research work to explore the potential of this language
is required to open perspectives in the computational lin-
guistic domain. Usage of Sanskrit becomes a challenging
task for MTS because of the morphological complexity
of the language. Currently, there are a few systems for
English to Sanskrit translation. However, more work in
this field is highly desirable [11] because, in India, most
of the holy books (Granths) are available in Sanskrit.
Even Sanskrit derives almost all Indic language.
• Usage of Ensemble techniques for developing multiple
systems requires manual intervention [69]. The merging
of multiple alternative generative systems and combining
outputs of different systems. The ensemble technique can
be applied using checkpoints, averaging the outputs and
re-ranking the translation from right to left in decoding.
• Dealing with rear words or unknown words in the vocab-
ulary of language pairs is tedious in building MTS [178].
Managing inventory and company names containing spe-
cial characters is also challenging. These words need to
be handled by forming a special entry as UNK tag or
entry in the back-off dictionary in machine-engineered
approaches. Whereas, in the case of human-engineering
these approaches are handled by manually constructing
rules.
• In most of the Indic languages, monolingual corpora is
present. Due to scarce parallel corpora, the machine-
engineered techniques are not applicable. Therefore,
there is an absolute need to either embed language
models in training the system or create synthetic par-
allel corpora to build MTS for Indic languages [179].
• Recent trend of deep learning in computer vision and
speech recognization has inspired the MT field to
develop deeper models. It reduces the model perplex-
ity, increases the accuracy and builds a better coverage
model [180].
• The alignment of the input sentence with the translated
sentence requires human intervention [181]. In human-
engineered approach manual rules for Vibhakti, TAM
and gerund handle alignment. Whereas in the case of
machine-engineered approach there are several tools such
as GIZA++ [149], fast-align [182] and attention mecha-
nism [69]. In computer-aided translations requires the
origin of output translation and optimized alignment.
• The machine-engineered approach sometime translates
some word multiple times while skips some word lead-
ing to over-generation and under-generation respectively.
Some possible solution is attention mechanism, feature
engineering and fertility of words [183].
• MT data may contain text from a different topic, style and
formality and so on. The convergence of text needs to be
handled by adopting the system with out-of-domain data
[184].
• Indic language has rich linguistic features loaded with
morphology, inflexion and so on. These features need to
be integrated into MTS to yield an efficient translation.
It can be embedded to input sentences, output sentences
and to build linguistically structured models [185].
• The languages with rich morphology are required to
come up with an entirely different approach in dealing
with rich linguistic words. As it is necessary to find a
way to scale up training a neural network both in terms of
computation and memory so that much larger vocabular-
ies for both source and target languages can be used [66].
• In machine-engineered approaches, performance
degrades on inflectional words and long sentences. To
handle this it requires linguistic knowledge of language
pairs. As suggested by [185] incorporating linguistic
information to train the corpus-based system improves
the system.
• Different neural architectures need to be explored, espe-
cially for the decoder [66].
• The technique for paraphrase-creation needs to be
explored from manual word substitutions to pivot tech-
nique of other translation systems. Future work possi-
bilities include handling of noisy paraphrase, evaluation
strategy, named entity, genre mismatch [186].
• Usage of Sanskrit becomes a challenging task for MTS.
There are some limited systems for English to Sanskrit
translation. However, more work in this field is highly
desirable [11].
• The Sanskrit grammar is well structured and least ambig-
uous. MTS in Sanskrit is not an easy task. For enhancing
MTS development source or target language, MTS is in
developing stages. There are some systems which are
confined to specific domains and have concise sentences
and phrases. Due to the morphological richness of, a
13

separate lexicon for Sanskrit sentence with morphologi-
cal details may be maintained in a database stored in the
form of the logic of a programming language [171].
• The technique used for Sanskrit to Hindi MTS is dic-
tionary-based (word to word translation), rule-based
and statistical-based. Further extensions can be made by
using other methods (example-based, corpus-based or a
hybrid technique) for Sanskrit to Hindi MTS.
• Decades of work of digitisation of Sanskrit text, lexi-
cal resources, and development of Sanskrit linguistic
resources have culminated in a collaborative effort to
develop an efficient MTS for Sanskrit to Hindi [172].
• Exploring the use of high-order statistical methods for
processing of Sanskrit language is desirable [187] as
current methodology lack generalisation, domain adapt-
ability and extension. Much work in the field of MTS for
low-resource language pairs and morphologically rich
languages such as Indic Languages is highly desirable
[176].
• Handle the most common problem of MTS, i.e., updates
by making it function as a repository. It can also be
formed as a virtual appliance to solve this problem. The
key benefit of developing virtual machine is fine granu-
larity while reducing the time for adding and removing
computational resources. Virtualisation also increases
the mobility of application and reduces deployment time.
The virtualisation can be deployed both on cloud and
standalone system [176].
9 Conclusion
In this paper, the review of different modelling technique
along with the research challenges is presented. It serves the
developers with resources required for modelling different
techniques such as corpus, domains, toolkits, techniques,
models, features and their evaluation measures. A compari-
son of research work on different Indic language pairs based
on modelling techniques is performed. It influences the work
on Sanskrit–Hindi language pair which is minimal, despite
holding an ancient scientific and comprehensive literature of
India. As per the analysis, the availability of the translation
systems for translating any language to English and Hindi
is more as compared to other languages. As per the analy-
sis, it can be concluded that the use of SMT based MTS is
more, i.e. 41%, as compared to others whereas the use of
hybrid MTS is 14%, rule-based, is 22%, and neural-based
is 23%. As per the analysis, Neural and hybrid approaches
perform better as compared to other techniques so, these
techniques may be considered for future use. In this paper
existing work for the Sanskrit language has been reviewed
and analyzed that most frequent approach i.e RBMT and
SMT were used with low accuracy. Therefore more accurate
13
M. Singh et al.
and efficient technique i.e, NMT and HBMT are required
to be implemented. The paper further contributes to open
issues, technical and linguistic challenges along with future
research directions in the field of MT for processing the San-
skrit language.
Compliance with Ethical Standards 
Conflict of interest  The authors declare that they have no conflict of
interest.
References
1. Trujillo A (2012) Translation engines: techniques for machine
translation. Springer, Berlin
2. Bharati A, Chaitanya V, Sangal R, Ramakrishnamacharyulu K
(1995) Natural language processing: a Paninian perspective.
Prentice-Hall, New Delhi
3. Chowdhury GG (2003) Natural language processing. Annu Rev
Inf Sci Technol 37(1):51–89
4. Drummer A (1996) Literature review: MT. MT for South African
Languages. http://peopl​e.cs.uct.ac.za/~bshar​wood/downl​oads.
html. Retrieved 28 May 2019
5. Hutchins J (2005) Current commercial machine translation sys-
tems and computer-based translation tools: system types and
their uses. Int J Transl 17(1–2):5–38
6. Hutchins J (2009) Multiple uses of machine translation and com-
puterised translation tools. Mach Transl 13–20
7. Bharati A, Kulkarni A (2008) Information coding in a language:
some insights from pan.inian grammar
8. Briggs R (1985) Knowledge representation in Sanskrit and arti-
ficial intelligence. AI Mag 6(1):32
9. Kulkarni A, Das M (2012) Discourse analysis of Sanskrit texts.
In: Proceedings of the workshop on advances in discourse analy-
sis and its computational aspects, pp 1–16
10. Mane D, Hirve A (2013) Study of various approaches in
machine translation for Sanskrit language. Int J Adv Res Technol
2(4):383–387
11. Raulji JK, Saini JR (2016) Sanskrit machine translation systems:
a comparative analysis. Int J Comput Appl 136(1):1–4
12. Antony P (2013) Machine translation approaches and survey
for Indian languages. Int J Comput Linguist Chin Lang Process
18(1):13–20
13. Garje G, Kharate G (2013) Survey of machine translation sys-
tems in India. Int J Nat Lang Comput 2(4):47–65
14. Saini S, Sahula V (2015) A survey of machine translation tech-
niques and systems for Indian languages. In: 2015 IEEE interna-
tional conference on computational intelligence & communica-
tion technology. IEEE, pp 676–681
15. Naskar S, Bandyopadhyay S (2005) Use of machine translation
in India: current status. AAMT J 36:25–31
16. Dwivedi SK, Sukhadeve PP (2010) Machine translation system
in Indian perspectives. J Comput Sci 6(10):1111
17. Hutchins WJ (2001) Machine translation over fifty years. Histoire
epistémologie langage 23(1):7–31
18. Hutchins WJ, Somers HL (1992) An introduction to machine
translation, vol 362. Academic Press, London
19. Noone G (2003) Machine translation a transfer approach. Com-
puter Science, Linguistics and a Language (CSLL) Department,
University of Dublin, Trinity College, Final Report
Machine Translation Systems for Indian Languages: Review of Modelling Techniques, Challenges,…
20. Dorr BJ, Hovy EH, Levin LS (2004) Machine translation: inter-
lingual methods. In: Natural language processing and machine
translation encyclopedia of language and linguistics, 2nd edn.
Elsevier, Amsterdam
21. Kavirajan B, Kumar MA, Soman K, Rajendran S, Vaithehi S
(2017) Improving the rule based machine translation system
using sentence simplification (English to Tamil). In: 2017 inter-
national conference on advances in computing. Communications
and informatics (ICACCI). IEEE, pp 957–963
22. Rana M, Atique M (2016) Use of fuzzy tool for example based
machine translation. Procedia Comput Sci 79:199–206
23. Darbari H, Kumar A, Dasgupta A, Mishra SK (2010) Complexity
of language in Rajya Sabha domain and MANTRA approach.
In: Proceedings of ICON-2010: 8th international conference on
natural language processing
24. Garje G, Kharate G, Kulkarni H (2014) Transmuter: an approach
to rule-based English to Marathi machine translation. Int J Com-
put Appl 98(21):33–37
25. Dubey P (2014) Need for Hindi–Dogri machine translation sys-
tem. In: 2014 international conference on computing for sustain-
able global development (INDIACom). IEEE, pp 136–140
26. Basavaraddi MCCS, Shashirekha DH (2014) A typical machine
translation system for English to Kannada. Int J Sci Eng Res 5(4).
ISSN 2229-5518
27. Adapanawar A, Garje A, Thakare P, Gundawar P, Kulkarni P
(2013) Rule based English to Marathi translation of assertive
sentence. Int J Sci Eng Res 4(5):516–518
28. Pishartoy D, Priya SW (2012) Extending capabilities of English
to Marathi machine translator. I. JCSI Int J Comput Sci Issues
9(3):375
29. Nair LR, Peter D, Ravindran RP (2012) Design and development
of a Malayalam to English translator—a transfer based approach.
Int J Comput Linguist (IJCL) 3(1):1–11
30. Batra KK, Lehal G (2010) Rule based machine translation of
noun phrases from Punjabi to English. Int J Comput Sci Issues
(IJCSI) 7(5):409
31. Rajan R, Sivan R, Ravindran R, Soman K (2009) Rule based
machine translation from English to Malayalam. In: International
conference on advances in computing, control, & telecommuni-
cation technologies, 2009, ACT’09. IEEE, pp 439–441
32. Sinha RMK, Mahesh K (2009) Developing English–Urdu
machine translation via Hindi. In: Third workshop on computa-
tional approaches to Arabic-script-based languages, Citeseer
33. Sinha R, Jain A (2003) AnglaHindi: an English to Hindi machine-
aided translation system. MT Summit IX, New Orleans, USA pp
494–497
34. Dave S, Parikh J, Bhattacharyya P (2001) Interlingua-based Eng-
lish–Hindi machine translation and language divergence. Mach
Transl 16(4):251–304
35. Kumar MA, Premjith B, Shivkaran S, Kavirajan B, Rajendran A,
Soman K (2017) Overview of the shared task on machine transla-
tion in Indian languages (MTIL-2017). J Intell Syst 28:455–464
36. Adak C (2014) A bilingual machine translation system: English
& Bengali. In: 2014 first international conference on automation,
control, energy and systems (ACES). IEEE, pp 1–4
37. Weaver W (1955) Translation. Mach Transl Lang 14:15–23
38. Brown PF, Pietra VJD, Pietra SAD, Mercer RL (1993) The math-
ematics of statistical machine translation: parameter estimation.
Comput Linguist 19(2):263–311
39. Vogel S, Ney H, Tillmann C (1996) HMM-based word alignment
in statistical translation. In: Proceedings of the 16th conference
on computational linguistics, vol 2. Association for Computa-
tional Linguistics, pp 836–841
40. Och FJ, Ney H (2003) A systematic comparison of various sta-
tistical alignment models. Comput Linguist 29(1):19–51
41. Koehn P, Och FJ, Marcu D (2003) Statistical phrase-based
translation. In: Proceedings of the 2003 conference of the North
American chapter of the association for computational linguistics
on human language technology, vol 1. Association for Computa-
tional Linguistics, pp 48–54
42. Yamada K, Knight K (2001) A syntax-based statistical transla-
tion model. In: Proceedings of the 39th annual meeting of the
Association for Computational Linguistics
43. Charniak E, Knight K, Yamada K (2003) Syntax-based language
models for statistical machine translation. In: Proceedings of MT
Summit IX, pp 40–46
44. Chiang D (2005) A hierarchical phrase-based model for statistical
machine translation. In: Proceedings of the 43rd annual meeting
on Association for Computational Linguistics. Association for
Computational Linguistics, pp 263–270
45. Subalalitha BSB Aarthi Venkataraman (2018) Statistical machine
translation system from English–Hindi. Int J Pure Appl Math
118(20):1649–1655
46. Jindal S, Goyal V, Bhullar JS (2018) English to Punjabi statistical
machine translation using moses (corpus based). J Stat Manag
Syst 21(4):553–560
47. Patel RN, Pimpale PB, Sasikumar M (2018) Machine translation
in Indian languages: challenges and resolution. J Intell Syst. https​
://doi.org/10.1515/jisys​-2018-0014
48. Khan NJ, Anwar W, Durrani N (2017) Machine translation
approaches and survey for Indian languages. arXiv preprint arXiv​
:17010​4290
49. Patel RN, Pimpale PB, Sasikumar M (2016) Statistical machine
translation for Indian languages: mission Hindi. arXiv preprint
arXiv​:16100​7418
50. Patel RN, Pimpale PB (2016) Statistical machine translation for
Indian languages: mission Hindi 2. arXiv​:16100​7418v​1
51. Das P, Baruah KK (2014) Assamese to English statistical
machine translation integrated with a transliteration module. Int
J Comput Appl 100(5):401–406
52. Ali A, Siddiq S, Malik MK (2010) Development of parallel cor-
pus and English to Urdu statistical machine translation. Int J Eng
Technol IJET-IJENS 10:31–33
53. Khan N, Anwar MW, Bajwa UI, Durrani N (2013) English to
Urdu hierarchical phrase-based statistical machine translation. In:
Proceedings of the 4th workshop on South and Southeast Asian
natural language processing, pp 72–76
54. Ali A, Hussain A, Malik MK (2013) Model for English–Urdu
statistical machine translation. World Appl Sci 24:1362–1367
55. Kumar P, Kumar V (2013) Statistical machine translation based
Punjabi to English transliteration system for proper nouns. Int J
Appl Innov Eng Manag 2(8):318–321
56. Anwar MM, Anwar MZ, Bhuiyan MAA (2009) Syntax analysis
and machine translation of Bangla sentences. Int J Comput Sci
Netw Secur 9(8):317–326
57. Udupa R, Faruquie TA (2004) An English–Hindi statistical
machine translation system. In: International conference on
natural language processing. Springer, pp 254–262
58. Cherry C (2008) Cohesive phrase-based decoding for statistical
machine translation. In: Proceedings of ACL-08: HLT, pp 72–80
59. Sutskever I, Vinyals O, Le QV (2014) Sequence to sequence
learning with neural networks. In: Advances in neural informa-
tion processing systems, pp 3104–3112
60. Cho K, Van Merriënboer B, Gulcehre C, Bahdanau D, Bougares
F, Schwenk H, Bengio Y (2014) Learning phrase representations
using RNN encoder–decoder for statistical machine translation.
arXiv preprint arXiv​:14061​078
61. Luong MT, Le QV, Sutskever I, Vinyals O, Kaiser L (2015)
Multi-task sequence to sequence learning. arXiv preprint arXiv​
:15110​6114
13

62. Jean S, Firat O, Cho K, Memisevic R, Bengio Y (2015) Montreal
neural machine translation systems for WMT’15. In: Proceed-
ings of the tenth workshop on statistical machine translation, pp
134–140
63. Bahdanau D, Cho K, Bengio Y (2014) Neural machine trans-
lation by jointly learning to align and translate. arXiv preprint
arXiv​:14090​473
64. Kalchbrenner N, Blunsom P (2013) Recurrent continuous transla-
tion models. In: Proceedings of the 2013 conference on empirical
methods in natural language processing, pp 1700–1709
65. Luong MT, Sutskever I, Le QV, Vinyals O, Zaremba W (2014)
Addressing the rare word problem in neural machine translation.
arXiv preprint arXiv​:14108​206
66. Cho K, Van Merriënboer B, Bahdanau D, Bengio Y (2014) On
the properties of neural machine translation: encoder–decoder
approaches. arXiv preprint arXiv​:14091​259
67. Chorowski JK, Bahdanau D, Serdyuk D, Cho K, Bengio Y (2015)
Attention-based models for speech recognition. In: Advances in
neural information processing systems. MIT Press, pp 577–585
68. Hochreiter S, Schmidhuber J (1997) Long short-term memory.
Neural Comput 9(8):1735–1780
69. Luong MT, Pham H, Manning CD (2015) Effective approaches to
attention-based neural machine translation. arXiv preprint arXiv​
:15080​4025
70. Meng F, Lu Z, Wang M, Li H, Jiang W, Liu Q (2015) Encoding
source language with convolutional neural network for machine
translation. arXiv preprint arXiv​:15030​1838
71. Kaiser Ł, Bengio S (2016) Can active memory replace atten-
tion? In: Advances in neural information processing systems.
MIT Press, pp 3781–3789
72. Kalchbrenner N, Espeholt L, Simonyan K, Oord Avd, Graves A,
Kavukcuoglu K (2016) Neural machine translation in linear time.
arXiv preprint arXiv​:16101​0099
73. Kaiser L, Gomez AN, Chollet F (2017) Depthwise separable
convolutions for neural machine translation. arXiv preprint arXiv​
:17060​3059
74. Oord Avd, Dieleman S, Zen H, Simonyan K, Vinyals O, Graves
A, Kalchbrenner N, Senior A, Kavukcuoglu K (2016) Wavenet:
a generative model for raw audio. arXiv preprint arXiv​:16090​
3499
75. Gehring J, Auli M, Grangier D, Yarats D, Dauphin YN (2017)
Convolutional sequence to sequence learning. Corr arXiv​
:1705.03122​
76. Jha S, Sudhakar A, Singh AK (2018) Neural machine translation
based word transduction mechanisms for low-resource languages.
arXiv preprint arXiv​:18110​8816
77. Choudhary H, Pathak AK, Saha RR, Kumaraguru P (2018) Neu-
ral machine translation for English–Tamil. In: Proceedings of the
third conference on machine translation: shared task papers, pp
770–775
78. Pathak A, Pakray P (2017) Neural machine translation for Indian
languages. J Intell Syst 28:465–477
79. Ramesh SH, Sankaranarayanan KP (2018) Neural machine trans-
lation for low resource languages using bilingual lexicon induced
from comparable corpora. In: Proceedings of the 2018 confer-
ence of the North American chapter of the association for com-
putational linguistics: student research workshop, pp 112–119
80. Singh S, Anand Kumar M, Soman K (2018) Attention based
English to Punjabi neural machine translation. J Intell Fuzzy Syst
34(3):1551–1559
81. Mistry J, Verma AA, Bhattacharyya P (2017) Literature survey:
study of neural machine translation. www.cfilti​ itbac​ in/resour​ ces/
surve​ys/ajay-jigar​-nmt-surve​y-jun17​pdf
13
M. Singh et al.
82. Revanuru K, Turlapaty K, Rao S (2017) Neural machine trans-
lation of Indian languages. In: Proceedings of the 10th annual
ACM India compute conference on ZZZ. ACM, pp 11–20
83. Tennage P, Sandaruwan P, Thilakarathne M, Herath A, Ranat-
hunga S, Jayasena S, Dias G (2017) Neural machine translation
for Sinhala and Tamil languages. In: 2017 international confer-
ence on Asian language processing (IALP). IEEE, pp 189–192
84. Agrawal R, Sharma DM (2017) Building an effective mt system
for English–Hindi using RNN’s. Int J Artif Intell Appl (IJAIA)
8(5):602–609
85. Das A, Yerra P, Kumar K, Sarkar S (2016) A study of attention-
based neural machine translation model on Indian languages. In:
Proceedings of the 6th workshop on South and Southeast Asian
natural language processing (WSSANLP2016), pp 163–172
86. Habash N, Dorr B, Monz C (2009) Symbolic-to-statistical
hybridization: extending generation-heavy machine translation.
Mach Transl 23(1):23–63
87. Sánchez-Martínez F, Forcada ML, Way A et al (2009) Hybrid
rule-based-example-based MT: feeding apertium with sub-sen-
tential translation units. In: Proceedings of the 3rd workshop on
example based machine translation, pp 11–18
88. Antonova A, Misyurev A (2014) Improving the precision of auto-
matically constructed human-oriented translation dictionaries.
In: Proceedings of the 3rd workshop on hybrid approaches to
machine translation (HyTra), pp 58–66
89. Göhring A (2014) Building a Spanish–German dictionary for
hybrid MT. In: Proceedings of the 3rd workshop on hybrid
approaches to machine translation (HyTra), pp 30–35
90. Sánchez-Martínez F, Forcada ML (2009) Inferring shallow-
transfer machine translation rules from small parallel corpora.
J Artif Intell Res 34:605–635
91. Tyers FM, Sánchez-Martínez F, Forcada ML et al (2012) Flexible
finite-state lexical selection for rule-based machine translation.
In: Proceedings of the 16th EAMT conference, Trento, Italy
92. Rudnick A, Gasser M (2013) Lexical selection for hybrid Mt with
sequence labeling. In: Proceedings of the second workshop on
hybrid approaches to translation, pp 102–108
93. Ruiz Costa-Jussà M, Centelles J (2015) Description of the Chi-
nese-to-Spanish rule-based machine translation system devel-
oped with a hybrid combination of human annotation and statisti-
cal techniques. ACM Trans Asian Lang Inf Process 15(1):1–13
94. Federmann C, Hunsicker S (2011) Stochastic parse tree selec-
tion for an existing RBMT system. In: Proceedings of the sixth
workshop on statistical machine translation. Association for
Computational Linguistics, pp 351–357
95. Dove C, Loskutova O, de la Fuente R (2012) What’s your pick:
RBMT, SMT or hybrid. In: Proceedings of the tenth conference
of the Association for Machine Translation in the Americas
(AMTA 2012), San Diego, CA
96. Hunsicker S, Yu C, Federmann C (2012) Machine learning for
hybrid machine translation. In: Proceedings of the seventh work-
shop on statistical machine translation. Association for Compu-
tational Linguistics, pp 312–316
97. Labaka G, España-Bonet C, Màrquez L, Sarasola K (2014) A
hybrid machine translation architecture guided by syntax. Mach
Transl 28(2):91–125
98. Crego J (2014) Systran RBMT engine: hybridization experi-
ments. In: 3rd workshop on hybrid approaches to machine trans-
lation (HyTra), Gothenburg, Sweden
99. Eberle K (2014) Hybrid strategies for better products and shorter
time-to-market. In: Proceedings of the 3rd workshop on hybrid
approaches to machine translation (HyTra), p 97
100. Simard M, Ueffing N, Isabelle P, Kuhn R (2007) Rule-based
translation with statistical phrase-based post-editing. In:
Machine Translation Systems for Indian Languages: Review of Modelling Techniques, Challenges,…
Proceedings of the second workshop on statistical machine trans-
lation. Association for Computational Linguistics, pp 203–206
101. Lagarda AL, Alabau V, Casacuberta F, Silva R, Díaz-de Liaño E
(2009) Statistical post-editing of a rule-based machine transla-
tion system. In: Proceedings of human language technologies:
the 2009 annual conference of the North American chapter of
the Association for Computational Linguistics, Companion Vol-
ume: Short Papers, Association for Computational Linguistics,
pp 217–220
102. Suzuki H (2011) Automatic post-editing based on SMT and its
selective application by sentence-level automatic quality evalu-
ation. Language 1:59–429
103. Béchara H, Rubino R, He Y, Ma Y, van Genabith J (2012) An
evaluation of statistical post-editing systems applied to RBMT
and SMT systems. In: Proceedings of COLING 2012, pp
215–230
104. Xia F, McCord M (2004) Improving a statistical MT system
with automatically learned rewrite patterns. In: Proceedings of
the 20th international conference on computational linguistics.
Association for Computational Linguistics, p 508
105. Collins M, Koehn P, Kučerová I (2005) Clause restructuring for
statistical machine translation. In: Proceedings of the 43rd annual
meeting on association for computational linguistics. Association
for Computational Linguistics, pp 531–540
106. Wang C, Collins M, Koehn P (2007) Chinese syntactic reor-
dering for statistical machine translation. In: Proceedings of
the 2007 joint conference on empirical methods in natural lan-
guage processing and computational natural language learning
(EMNLP-CoNLL)
107. Patel RN, Gupta R, Pimpale PB et al (2016) Reordering rules for
English–Hindi SMT. arXiv preprint arXiv​:16100​7420
108. Farrús M, Costa-Jussa MR, Marino JB, Poch M, Hernández
A, Henríquez C, Fonollosa JA (2011) Overcoming statistical
machine translation limitations: error analysis and proposed
solutions for the Catalan–Spanish language pair. Lang Resour
Eval 45(2):181–208
109. Formiga Fanals L, Hernández Huerta A, Mariño Acebal JB,
Monte Moreno E (2012) Improving English to Spanish out-
of-domain translations by morphology generalization and gen-
eration. In: Proceedings of the monolingual machine transla-
tion-2012 workshop, pp 6–16
110. Carl M, Pease C, Iomdin LL, Streiter O (2000) Towards a
dynamic linkage of example-based and rule-based machine
translation. Mach Transl 15(3):223–257
111. Hua W, Haifeng W (2004) Improving statistical word alignment
with a rule-based machine translation system. In: Proceedings of
the 20th international conference on computational linguistics.
Association for Computational Linguistics, p 29
112. Okuma H, Yamamoto H, Sumita E (2008) Introducing a transla-
tion dictionary into phrase-based SMT. IEICE Trans Inf Syst
91(7):2051–2057
113. Eisele A, Federmann C, Saint-Amand H, Jellinghaus M, Her-
rmann T, Chen Y (2008) Using Moses to integrate multiple
rule-based machine translation engines into a hybrid system. In:
Proceedings of the third workshop on statistical machine transla-
tion, pp 179–182
114. Sánchez-Cartagena VM, Sánchez-Martínez F, Pérez-Ortiz JA
et al (2011) Integrating shallow-transfer rules into phrase-based
statistical machine translation. In: Machine translation summit.
pp 562–569
115. Chen Y, Eisele A (2010) Integrating a rule-based with a hierar-
chical translation system. In: LREC
116. Ahsan A, Kolachina P, Kolachina S, Sharma DM, Sangal R
(2010) Coupling statistical machine translation with rule-based
transfer and generation. In: Proceedings of the 9th conference
of the association for machine translation in the Americas
117. Enache R, España  Bonet C, Ranta A, Màrquez  Villodre L
(2012) A hybrid system for patent translation. In: Proceedings
of the 16th annual conference of the European Association for
Machine Translation: EAMT 2012: Trento, Italy, May 28th–
30th 2012, pp 269–278
118. Och FJ, Ney H (2004) The alignment template approach to sta-
tistical machine translation. Comput Linguist 30(4):417–449
119. Groves D, Way A (2005) Hybrid data-driven models of
machine translation. Mach Transl 19(3–4):301–323
120. Wang K, Zong C, Su KY (2013) Integrating translation mem-
ory into phrase-based machine translation during decoding.
In: Proceedings of the 51st annual meeting of the Association
for Computational Linguistics (vol 1: Long Papers), vol 1, pp
11–21
121. Carbonell JG, Klein S, Miller D, Steinbaum M, Grassiany T, Frey
J (2006) Context-based machine translation. In: 7th conference
of the Association for Machine Translation in the Americas
122. Vandeghinste V, Schuurman I, Carl M, Markantonatou S, Badia
T (2006) METIS-II: machine translation for low resource lan-
guages. In: LREC, pp 1284–1289
123. Ruiz Costa-Jussà M, Rodríguez  Fonollosa JA (2011) Using
linear interpolation and weighted reordering hypotheses in the
Moses system. In: Seventh conference on international language
resources and evaluation, pp 1712–1718
124. Tambouratzis G, Sofianopoulos S, Vassiliou M (2013) Language-
independent hybrid MT with present. In: Proceedings of the sec-
ond workshop on hybrid approaches to translation, pp 123–130
125. Dhariya O, Malviya S, Tiwary US (2017) A hybrid approach for
Hindi–English machine translation. In: 2017 international con-
ference on information networking (ICOIN). IEEE, pp 389–394
126. Salunkhe P, Kadam AD, Joshi S, Patil S, Thakore D, Jadhav
S (2016) Hybrid machine translation for English to Marathi: a
research evaluation in machine translation: (hybrid translator).
In: International conference on electrical, electronics, and opti-
mization techniques (ICEEOT). IEEE, pp 924–931
127. Nithya B, Joseph S (2013) A hybrid approach to English to
Malayalam machine translation. Int J Comput Appl 81(8):11–15
128. Kaur H, Laxmi DV (2013) A web based English to Punjabi MT
system for news headlines. Int J Adv Res Comput Sci Softw Eng
3(6):1092–1094
129. Dhore M, Dixit S, Karande J (2011) Web page interface locali-
sation in Devanagari for commercial interactive applications by
enhancing basic functionality of apache server. Int J Comput
Appl 18(4):6–10
130. Chatterji S, Sonare P, Sarkar S, Basu A (2011) Lattice based
lexical transfer in Bengali Hindi machine translation framework.
In: Proceedings of ICON-2011: 9th international conference on
natural language processing
131. Shahnawaz Mishra R (2015) An English to Urdu translation
model based on CBR, ANN and translation rules. Int J Adv Intell
Paradig 7(1):1–23
132. Chatterji S, Roy D, Sarkar S, Basu A (2009) A hybrid approach
for Bengali to Hindi machine translation. In: 7th international
conference on natural language processing, pp 83–91
133. Dorr BJ (1994) Machine translation divergences: a for-
mal description and proposed solution. Comput Linguist
20(4):597–633
134. Habash N, Dorr B (2002) Handling translation divergences:
combining statistical and symbolic techniques in generation-
heavy machine translation. In: Conference of the association for
machine translation in the Americas. Springer, pp 84–93
135. Goyal P, Sinha RMK (2009) Translation divergence in English–
Sanskrit–Hindi language pairs. In: International Sanskrit com-
putational linguistics symposium. Springer, pp 134–143
136. Mishra V, Mishra R (2008) Study of example based English to
Sanskrit machine translation. Polibits 37:43–54
13

137. Shukla P, Shukl D, Kulkarni A (2010) Vibhakti divergence
between Sanskrit and Hindi. In: International Sanskrit computa-
tional linguistics symposium. Springer, pp 198–208
138. Luong MT, Manning CD (2015) Stanford neural machine transla-
tion systems for spoken language domains. In: Proceedings of the
international workshop on spoken language translation, pp 76–79
139. Farajian MA, Turchi M, Negri M, Bertoldi N, Federico M (2017)
Neural vs. phrase-based machine translation in a multi-domain
scenario. In: Proceedings of the 15th conference of the European
chapter of the association for computational linguistics: volume
2, short papers, vol 2, pp 280–284
140. Turchi M, De Bie T, Cristianini N (2008) Learning performance
of a machine translation system: a statistical and computational
analysis. In: Proceedings of the third workshop on statistical
machine translation. Association for Computational Linguistics,
pp 35–43
141. Irvine A, Callison-Burch C (2013) Combining bilingual and
comparable corpora for low resource machine translation. In:
Proceedings of the eighth workshop on statistical machine trans-
lation, pp 262–270
142. Chen B, Kuhn R, Foster G, Cherry C, Huang F (2016) Bilingual
methods for adaptive training data selection for machine transla-
tion. In: Proceedings of AMTA, pp 93–103
143. Arthur P, Neubig G, Nakamura S (2016) Incorporating discrete
translation lexicons into neural machine translation. arXiv pre-
print arXiv​:16060​2006
144. Tu Z, Lu Z, Liu Y, Liu X, Li H (2016) Modeling coverage for
neural machine translation. arXiv preprint arXiv​:16010​4811
145. Chen W, Matusov E, Khadivi S, Peter JT (2016) Guided align-
ment training for topic-aware neural machine translation. arXiv
preprint arXiv​:16070​1628
146. Liu L, Utiyama M, Finch A, Sumita E (2016) Neural machine
translation with supervised attention. arXiv preprint arXiv:​ 16090​
4186
147. Liu Y, Liu Q, Lin S (2006) Tree-to-string alignment template for
statistical machine translation. In: Proceedings of the 21st inter-
national conference on computational linguistics and the 44th
annual meeting of the association for computational linguistics.
Association for Computational Linguistics, pp 609–616
148. Och FJ, Ney H (2000) A comparison of alignment models for
statistical machine translation. In: COLING 2000, the 18th inter-
national conference on computational linguistics, vol 2
149. Tian L, Wong F, Chao S (2011) Word alignment using Giza++
on windows. Mach Transl 2:1762–1765
150. Deng Y, Byrne W (2006) MTTK: an alignment toolkit for statisti-
cal machine translation. In: Proceedings of the 2006 conference
of the North American chapter of the association for computa-
tional linguistics on human language technology: companion vol-
ume: demonstrations. Association for Computational Linguistics,
pp 265–268
151. Ortiz-Martínez D, García-Varea I, Casacuberta F (2005) THOT:
a toolkit to train phrase-based statistical translation models. In:
Tenth machine translation, Citeseer
152. Koehn P, Knowles R (2017) Six challenges for neural machine
translation. arXiv preprint arXiv​:17060​3872
153. Koehn P (2009) Statistical machine translation. Cambridge Uni-
versity Press, Cambridge
154. Okpor M (2014) Machine translation approaches: issues and
challenges. Int J Comput Sci Issues (IJCSI) 11(5):159
155. Koehn P, Axelrod A, Mayne AB, Callison-Burch C, Osborne M,
Talbot D, White M (2005) Edinburgh system description for the
2005 NIST MT evaluation. In: Proceedings of machine transla-
tion evaluation workshop 2005
156. Pingali P, Varma V (2006) Hindi and Telugu to English cross
language information retrieval at CLEF 2006. In: CLEF (work-
ing notes)
13
M. Singh et al.
157. Dhore M, Dixit S (2011) English to Devanagari translation for
UI labels of commercial web based interactive applications.
Int J Comput Appl 35(10):6–12
158. Godase A, Govilkar S (2015) A novel approach for rule based
translation of English to Marathi. Adv Comput Intell Int J.
https​://doi.org/10.5121/acii.2015.2401
159. Savita Singla SB (2013) Hybrid approach for English to Pun-
jabi translation system for news paper headlines in a specific
domain. Int J Eng Res Technol 2:1792–1795
160. Sinhal RA, Gupta KO (2014) A pure EBMT approach for
English to Hindi sentence translation system. Int J Mod Educ
Comput Sci 6(7):1
161. Anju E, Manoj Kumar K (2014) Malayalam to English
machine translation: an EBMT system. IOSR J Eng (IOSRJEN)
4(01):18–23
162. Census of India Government (2018) Census-2018, “language-
census of India, states and union territories”. www.censu​sIndi​
a.gov.in/2011C​ensus​/C-16_25062​018_NEW.pdf. Accessed 29
May 2019
163. Bahadur P, Jain A, Chauhan D (2011) English to Sanskrit
machine translation. In: Proceedings of the international con-
ference & workshop on emerging trends in technology. ACM,
pp 641–645
164. Jha GN, Mishra SK, Chandrashekar R (2005) Developing a
Sanskrit analysis system for machine translation. In: Proceed-
ings of national seminar on translation today: state and issues.
Department of Linguistics, University of Kerala, Trivandrum,
pp 23–25
165. Bharati A, Kulkarni A (2007) Sanskrit and computational lin-
guistics. In: First international Sanskrit computational sympo-
sium. Department of Sanskrit Studies, University of Hyderabad
166. Huet G (2009) Formal structure of Sanskrit text: requirements
analysis for a mechanical Sanskrit processor. In: Nath Jha G (ed)
Sanskrit computational linguistics. Springer, Berlin, pp 162–199
167. Baindur M (2015) Nature in Indian philosophy and cultural tradi-
tions. Springer, Berlin
168. Bharati A, Chaitanya V, Sangal R (1994) Paninian framework
and its application to Anusaraka. Sadhana 19(1):113–127
169. Nair RR, Devi LS (2011) Sanskrit informatics: informatics for
Sanskrit studies and research. Centre for Informatics Research
and Development, Trivandrum
170. Pandey RK, Jha GN (2016) Error analysis of Sahit—a statistical
Sanskrit–Hindi translator. Procedia Comput Sci 96:495–501
171. Rathod SG (2014) Machine translation of natural language using
different approaches. Int J Comput Appl 102(15):26–31
172. Shukla P, Shukla A (2014) English speech to Sanskrit speech
(ESSS) using rule based translation. Int J Comput Appl
92(10):37–42
173. Bahadur P, Jain A, Chauhan D (2012) ETRANS—a complete
framework for English to Sanskrit machine translation. In: Inter-
national Journal of Advanced Computer Science and Applica-
tions (IJACSA) from international conference and workshop on
emerging trends in technology, Citeseer
174. Jayan V, Sunil R, Kurambath GS, Kumar RR (2012) Divergence
patterns in machine translation between Malayalam and English.
In: Proceedings of the international conference on advances in
computing, communications and informatics. ACM, pp 788–794
175. Aparna S (2005) Sanskrit to English translator. In: Language
in India, vol 5. http://www.langu​agein​india​.com/jan20​05/aparn​
asans​kritd​isser​tatio​n1.html. Accessed 29 May 2019
176. Lopez A, Post M (2013) Beyond bitext: five open problems in
machine translation. In: Proceedings of the EMNLP workshop
on twenty years of Bitext
177. Pathak KN, Jha GN (2011) Challenges in NP case-mapping in
Sanskrit Hindi machine translation. In: Information systems for
Indian languages. Springer, pp 289–293
Machine Translation Systems for Indian Languages: Review of Modelling Techniques, Challenges,…
178. Li X, Zhang J, Zong C (2016) Towards zero unknown word in
neural machine translation. In: IJCAI, pp 2852–2858
179. Sennrich R, Haddow B, Birch A (2015) Improving neural
machine translation models with monolingual data. arXiv pre-
print arXiv​:15110​6709
180. Britz D, Goldie A, Luong MT, Le Q (2017) Massive exploration
of neural machine translation architectures. arXiv preprint arXiv​
:17030​3906
181. Alkhouli T, Bretschner G, Peter JT, Hethnawi M, Guta A, Ney
H (2016) Alignment-based neural machine translation. In: Pro-
ceedings of the first conference on machine translation: volume
1, research papers, vol 1, pp 54–65
182. Lamraoui F, Langlais P (2013) Yet another fast, robust and open
source sentence aligner. Time to reconsider sentence alignment.
In: XIV machine translation summit
183. Takebayashi Y, Chenhui C, Arase Y, Nagata M (2018) Word
rewarding for adequate neural machine translation. In: Interna-
tional workshop on spoken language translation
184. Freitag M, Al-Onaizan Y (2016) Fast domain adaptation for neu-
ral machine translation. arXiv preprint arXiv​:16120​6897
185. Sennrich R, Haddow B (2016) Linguistic input features improve
neural machine translation. arXiv preprint arXiv​:16060​2892
186. Madnani N, Ayan NF, Resnik P, Dorr BJ (2007) Using para-
phrases for parameter tuning in statistical machine translation. In:
Proceedings of the second workshop on statistical machine trans-
lation. Association for Computational Linguistics, pp 120–127
187. Goyal P, Huet G, Kulkarni A, Scharf P, Bunker R (2012) A dis-
tributed platform for Sanskrit processing. In: Proceedings of
COLING 2012, pp 1011–1028
Publisher’s Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
13