International Journal on Natural Language Computing (IJNLC) Vol. 6, No.

3, June 2017

A HYBRID APPROACH USING PHRASES AND RULES

FOR HINDI TO ENGLISH MACHINE TRANSLATION

Susmita Gupta and Niladri Chatterjee

Department of Mathematics, IIT Delhi, Delhi, India

ABSTRACT
The present work focuses on developing a hybrid approach for developing a machine translation (MT)
scheme for automatic translation of Hindi sentences to English. Development of machine translation (MT)
systems for Indian languages to English almost invariably suffers from the limited availability of linguistic
resources. As a consequence, statistical, rule-based or example-based approaches have not been found to
generate good quality translation for all types of sentences. Moreover, purely statistical or example based
schemes do not leverage the semantic association (i.e. association of the noun case-endings (karakas)) with
each other and also with the verb phrase to generate the English translation.
The scheme proposed in this work is based on the syntactic and semantic roles of the phrases of an input
Hindi Sentence. The proposed solution is a hybrid scheme involving phrase-based, rule-based and
statistical translation approaches. In our approach we try to identify the difficulties associated with the
translation of different types of input sentences, viz. simple declarative, negative and interrogative
sentences and propose a translation approach for them.

KEYWORDS
Hybrid MT, Rule-based MT, Example-based MT, Statistical MT, Noun case-ending.

1. INTRODUCTION
Past works on machine translation for Indian languages had mostly focused on translating English
sentences to Indian languages. Not much work is found in the reverse direction i.e. from Indian
languages to English. The primary reason for this may be attributed to the lack of resources for
processing Indian languages. The situation becomes even more difficult because of the lack of
existing parallel corpus essential for different translation paradigms, such as, example based
machine translation (EBMT), statistical machine translation (SMT), among others. Furthermore,
only a handful of parallel corpora are available for Hindi to English as one may find in the OPUS1
corpus. Consequently, the quality of translation as obtained from various translators available
online is not quite up to an acceptable level even for rather simple sentences. Table 1 shows some
examples of translation results using Google translator which is one of the most popular online
translator available. The table contains the translation of the Hindi sentence to English and the
BLEU score representing the quality of the translation. The BLEU metric scores a translation on a
scale of 0 to 1, but is frequently displayed as a percentage value. The closer the BLEU score is to
100%, that means the translation is more correlated to a human translation. Technically, the
BLEU metric measures how many words overlap in a given translation when compared to a
reference translation, giving higher scores to sequential words. It can be seen in Table 1 that the
BLEU score obtained for Hindi to English Translator of Google translation are quite low, which
can be interpreted also by the quality of translated results to a linguist who has the knowledge of
both the languages. Note that the type of sentences been chosen for translation are of a special
type (multi-karaka). This means that there are multiple occurrences of noun-case endings

http://opus.lingfil.uu.se/
DOI: 10.5121/ijnlc.2017.6305 57
 International Journal on Natural Language Computing (IJNLC) Vol. 6, No.3, June 2017
(karakas) in each of the Hindi sentences. The details of noun-case endings (karakas) are
explained later in this section.

To address the challenges of Hindi to English machine translation and to build a good quality
translation system, the inherent differences of the sentence structure between these two languages
have to be carefully observed. The structures of English and Hindi sentences vary a lot.
Development of a translation system needs considering and remembering these differences for
producing a correct translation. The most important of them are mentioned below:

1. Word Order: English is classified as an SVO (Subject-Verb-Object) language and Hindi

as an SOV (Subject-Object-Verb) language. An SVO language indicates, the position
preceding a verb marks the subject and the object follows the verb. On the other hand,
Hindi is a relatively free word-order language. For example, the following English
sentence Ram(S) eats (V) fruits (O) is written as (raam(S) phal(O) khaataa
hai(V)). However, reordering of the constituent words often does not change the
semantics of the sentence. For illustration,

(phal khaataa hai raam)

(raam phal khaataa hai)

are also correct Hindi sentences conveying the same sense as the previously mentioned one [8]
However, an MT system often fails to capture this equivalence. For illustration consider
the Google translation of the two sentences

(chaawal khaataa hai raam),

(raam chaawal khaataa hai).

Although both convey the same sense which is Ram eats rice, Google translates them
as: Rice account is Ram and Ram eats rice, respectively.

2. Order of main verb and auxiliary verbs: The relative order of the main verb and the
auxiliary verb in a sentence are reversed in Hindi and English. The difference between
Hindi and English verb groups is the order of root verb and its auxiliaries. In Hindi the
main root verb is followed by its suffixes, as in (jaa rahaa hai ~ going) where jaa
rahaa hai is the present continuous form of the verb jaanaa ~ to go. Here rahaa hai is the
auxiliary verb that conveys the tense, gender and case information of the verb. However,
in English the main verb follows the auxiliary verbs as in is going.

Table 1. Translation of Hindi to English and vice versa

58
 International Journal on Natural Language Computing (IJNLC) Vol. 6, No.3, June 2017

3. Presence of subject in a sentence: The subject position in English cannot be empty and
this leads to a forceful introduction of dummy it or existential there in certain
sentences to fill the subject position. Hindi sentence construction does not need any such
syntactic requirement. For illustration, consider the sentence "It is raining" ~
(baarish ho rahii hai). The translation does not contain any word corresponding to 'It'.

4. Preposition vs. Postposition: In English, prepositions come prior to the words to which
they relate. In Hindi, such words occur after the nouns they govern. For example, on the
door translates to (darvaaze par). Here darvaaze implies door, and par
corresponds to the preposition on. As they occur after the noun, they are often referred
to as postpositions.

Based on the output from Table 1 and the differences of the two languages, Hindi and English,
the basic problems of translations between Hindi to English sentences may be identified as
follows:

1. Semantic information is not preserved in some of the translations. This is because based
on the context of the sentence, a Hindi verb can be translated in various ways.
For example: In the Hindi sentence
(phool khil rahaa hai)
[Flower is blooming],
(khil) [blooming] should be translated as blooming. But in the sentence

(yah poshaak aap par khil rahii hai)

[This dress suits you],
(khil) [blooming] should be translated as suits.

2. Syntactically some translations are incorrect. The reason for this can be because the
grammatical differences mentioned above is not considered in the statistical and the
dictionary based schemes of translation.
3. In some cases, a few source language words are not translated at all in the target
language. For example, the Google translator does not translate suiting at all and places
the word as is in the translation.

59
 International Journal on Natural Language Computing (IJNLC) Vol. 6, No.3, June 2017
The problems become even moreworse for the case when the source language (Hindi) has
multiple noun case-endings (karakas) scenarios. For example, consider the following two Hindi
sentences, which are structurally same but semantically different.

In the first sentence, the sense that the noun case-ending (kii) conveys is that Ram took a watch
"made of" gold, whereas in the second sentence (kii) means that the watch "belongs to" Shyam.
But for existing translators (eg Google) the statements fail to preserve the semantic information of
the sentences as follows:

1. Ram took a gold clock for me

2. Ram took a look at Shyam's clock for me

which is clearly not preserving the semantics for the second sentence.

It is observed that the translation quality of statistical techniques [17] suffers because of the
inherent property of the scheme, which is to generate the translation based on the probability
computation from existing parallel corpora. Statistical schemes typically do not consider the
syntactic or semantic information of the source and target languages. It is clear from the outputs
given in Table 1 that the translations are neither faithful to the source sentence semantics nor
fluent in the target language. More recently, syntax has been incorporated in SMT in both the
source and the target, e.g. Syntax Augmented Machine Translation (SAMT) [27], Hierarchical
Phrase-Based SMT (HPBSMT) [10]. Also, different phrase-based SMT schemes are increasingly
looking forward towards incorporating semantic information into the translation. For example,
Combinatory Categorial Grammar (CCG) augmented hierarchical phrase-based machine
translation [1], CCG supertags in factored statistical machine translation [7].
Rule-based schemes also do not fit well in the purview for translation of Indian languages to
English. This is because of the huge differences in the sentence structure of the two languages
which leads to the need of a large set of translation rules to be developed. The sentence structure
of Hindi as against English is free word order, which makes the task of translation rule creation a
huge human effort.
The translations in example-based [24] schemes depend solely on the examples in the available
corpus. A strong parallel corpus covering variety of sentence types is not widely available for
Indian languages. Also, divergence between the two languages results in the lesser availability of
useful translation examples even if similar structure input sentences appear in the corpus.
Divergence [14] occurs when structurally similar sentences of the source language do not
translate into sentences that are similar in structures in the target language. For illustration,
consider the translations of the following English sentences:

The above drawbacks motivated us towards developing a hybrid translation scheme that can take
the advantages of each of the above schemes, and generate improved Hindi to English translation
system.

The paper is organized as follows. Section 2 describes some of the previous works in the
direction of translation involving Indian languages. This includes the work done in Indian
60
 International Journal on Natural Language Computing (IJNLC) Vol. 6, No.3, June 2017
language translations for the case of multiple noun case-endings (karakas). The proposed
translation approach is explained in Section 3, which is followed by Section 4 explaining the
experiments conducted, and shares the translation results. Finally, Section 5 concludes the paper,
by proposing few areas of future work.

2. PREVIOUS WORK

Over the last two decades' machine translation using Indian languages made some significant
progress. Perhaps the most notable among them is the AnglaBharati system [21] which uses rules
developed based on structural patterns to translate English sentences into different Indian
languages. However, several other systems have been developed prior to it, following different
translation paradigms. The most notable among them are mentioned below.

Anusaaraka [5], a translation system for mutual translation between different Indian languages is
developed using the direct translation approach maintaining the grammatical structure of the
source language. Hence the output is often syntactically incorrect when the source and target
language grammars differ. The Mantra machine translation system 2 was developed using the
transfer based approach. Here Tree Adjoining Grammar based taggers and parsers are used for
transferring English text structure to Hindi to achieve translation. Anubharati [15] followed a
hybrid approach where example based scheme was aided by statistical analysis corpora and
linguistic rules. Another EBMT system Anubaad [3] also followed EBMT approach for
translating English news sentences into Hindi. More recently many other systems have been made
available for translating between various Indian languages. Some of them are: Punjabi to Hindi
MT system [16], Hindi to Punjabi MT [13], UNL-Based systems for translating from English to
Hindi, Bengali and Marathi [9], EBMT systems for translating between Hindi, Kannada and
Tamil [2]. Barring the Hinglish translation system [22] none of the above translation schemes
attempted to translate Hindi to English. The existing schemes focus mostly in translating English
to Indian language, or between pairs of Indian languages.

Very recently in WMT14 [12] shared task, attempt has been done to develop statistical systems
for Hindi to English, and English to Hindi translation. The core components of the translation
systems are Phrase Based (Hindi- English) and factor based (English to Hindi). The work focuses
on the usage of number, case, and Tree Adjoining Grammar information for translation. There is
also a pre and post processing step involved to adjust the translation output for structural
divergence between Hindi and English sentences. Also, there are a very limited set of rules which
play the role in translating Hindi to English. Another translation system was developed by [26] in
WMT2014 where apart from various other language pairs, Hindi - English translation has also
been considered. This is based on string to tree system, where some improvements were done for
out of vocabulary word translation through transliteration.

Apart from the WMT shared task translation systems [18] presented a hybrid machine translation
architecture guided by syntax. In this work, the authors develop a hybrid translation system
guided by the rule based machine translation engine and, before transference, few partial
candidate translations given by the statistical translation schemes. These partial candidate
translations are used to enrich the tree-based representation. The final hybrid translation is created
by choosing the most probable combination among the available subsets with a statistical
decoder.

1 https://mantra-rajbhasha.rb-aai.in/
61
 International Journal on Natural Language Computing (IJNLC) Vol. 6, No.3, June 2017
In the works described above, the syntactical and semantic information is not utilized completely
to generate the translations. None of the translation approaches provide a complete set of
translation rules guided by syntactical information. Further, there has not been any special focus
on translating negative, and interrogative sentences which clearly has a different sentence
structure, and therefore different translation rules apply for them as against simple positive
sentences.

The present work focuses on Hindi to English translation. In our approach we focus on the
structural/syntactic representation and semantic information to develop a strong translation
paradigm. The structural information is captured by considering the syntactic phrases as units for
translation. The semantic information obtained from the (karaka) (case-ending) is used to
generate the translation output. The details are explained in Section 3.

The evaluation of the translation output is done using the BLEU score. Although there are various
metrics currently available for evaluating machine translation [28], the current work uses BLEU,
as it is the most widely used evaluation metric till date. [29]

3. PROPOSED APPROACH

The translation approach followed in this work is a hybrid of phrase-based, rule-based and
statistical approaches of machine translation. The strengths of each of the following approaches
are used, to generate a hybrid scheme for machine translation.

3.1. Phrase-based Approach

This approach translates a given sentence phrase wise, and then recombines the individual
translated phrases to generate the complete translation of the whole sentence. In the present work
the phrases considered are the syntactic phrases derived from the LTRC Hindi Shallow Parser. 3

3.2 Rule-based Approach

The translation of each phrase is recombined based on certain rules, designed specifically to solve
the given sentence structure. The basis of these rules is derived from
The translation pattern followed for Hindi to English with respect to various (karaka)
(noun case-ending) present in the Hindi Sentence.
Type of the sentence (simple declarative, negative or interrogative)
Once these (karaka) (noun case-ending) and the sentence type are identified properly, the
translation of a Hindi Sentence can be achieved using some transfer rules explained in the Section
6.
3.3 Statistical Approach
Each phrase is translated using statistical techniques. The two well-known statistical translation
techniques used for the current work are:

3
http://ltrc.iiit.ac.in/analyzer/Hindi/
62
 International Journal on Natural Language Computing (IJNLC) Vol. 6, No.3, June 2017
3.3.1 MOSES

Moses4 is a statistical machine translation system. MOSES allows the language models for any
language pair to be trained using a collection of parallel corpus. A parallel corpus consists of a set
of translated texts of the two languages under consideration called the source and the target
language. Word alignment of the parallel data is established using GIZA++5. Using these inputs a
trained target language model is generated using a toolkit called SRILM6 (SRI Language
Modelling Toolkit), which ensures the fluency of the output. Using the language model generated
by SRILM, any search algorithm can be used for finding the highest probability translation from
among the number of choices available. One such efficient search algorithm used widely is the
beam search7.
3.3.2 Google translator

Google Translator is a utility provided by Google to translate written text from one language to
another. This is one of the most popular and well known translators available currently.
The overall translation flow proposed is shown in Figure 1. Each step of the flow is explained as
follows:
a) Parsing and Sentence Type identification of the Hindi sentence
The first step for translation of any Hindi sentence through the proposed scheme is parsing it.
With this the syntactic phrases and the sentence type are identified.
The parsing is done using the LTRC shallow parser:
1. To identify syntactic phrases from the given Hindi sentence.
2. To identify the sentence type; positive, negative, or interrogative.
For example, consider the sentence
(prithvii har 12 mahine mein suraj kaa ek chakkar lagaatii
hai)

~ The earth rotates around the sun once in every 12 months.

The parse output for the sentence is shown in Table 2. In Table 2, NP signifies the Noun Phrase,
and VGF signifies the finite verb phrase.
In the given example five syntactic phrases are identified which are
1. (prithvii) ~ (Earth)
2. (har
12 mahine mein) (in every 12 months)
3. (suraj kaa) (of the sun)
4. (ek chakkar) (one round)
5. (lagaatii hai) (takes)
The NEG (Negative) and WQ (Question Word) POS tags are used to signify negative and
interrogative sentences.
Example of the parse output of a negative Hindi sentence

4
http://www.statmt.org/moses/
5
https://code.google.com/p/giza-pp/
6
http://www.speech.sri.com/cgi-bin/run-distill?papers/icslp2002-srilm.ps.gz
7
http://en.wikipedia.org/wiki/Beam_search

63
 International Journal on Natural Language Computing (IJNLC) Vol. 6, No.3, June 2017
(Raam khaanaa nahii khaataa hai) ~ Ram does not eat food

and Interrogative Hindi sentence

(Raam ne kyaa khaayaa) ~ What did Ram eat
is given in Table 3 and Table 4, respectively.

Input Hindi Sentence

Parsing and identification of the Sentence Type

of the Hindi sentence using shallow parser

Noun Case-ending identification

Phrase Generation

Phrase Translation

Rule based Phrase Recombination

Output English
Sentence
Figure 1 Overall Translation Flow
Table 2 Parse output from the LTRC Shallow Parser

64
 International Journal on Natural Language Computing (IJNLC) Vol. 6, No.3, June 2017

Table 3 Parse output for Negative Hindi Sentence

Table 4 Parse output for Interrogative Hindi Sentence

b) Noun Case-ending identification

Once the individual syntactic phrases are identified, the karaka (noun-case ending) information
associated to each Noun Phrase is to be identified. This is done using the syntactic cues for
identifying each karaka, defined in AnnCorra: TreeBanks for Indian Languages Guidelines for
Annotating Hindi TreeBank [4]. For example, the syntactic cues for identifying kartaa as mentioned
in [4] are as follows:
(a) Kartaa is normally in nominative case which is realized as in Hindi.
(b) By default verb in active voice agrees with the karta in number, gender, and person.

For the example sentence above, the karaka information for each Noun Phrase is as follows:

c) Phrase generation
The final phrases for translation are generated by the following two rules:

65
 International Journal on Natural Language Computing (IJNLC) Vol. 6, No.3, June 2017
1. Merge the sambandh (relation) noun case ending with the Noun Phrase (NP) following it.
This is done to preserve the inter noun phrase associations after the translation which are
depicted by the relation case ending words kaa/ke/kii/raa/re/rii.

2. For example, in the above phrases, (suraj kaa) is the sambandh (relation) noun case
ending. Hence, it will merge with the noun phrase following it. Hence, the final phrases
to be considered by the proposed scheme are:

3. If the sentence to be translates is identified as a negative sentence, then the negative

(NEG) word is added to the verb phrase in the Phrase Generation step. This is done
because the negation is associated with the Verb Phrase of the sentence.

For example, the sentence has the NEG word (nahii)

d) Phrase translation

The next step of the proposed scheme is to translate each phrase individually through statistical
techniques. As mentioned in the previous sections, two well-known statistical schemes are used
for the current work.
MOSES decoder, with GIZA++ and SRILM
If a good training corpus is available for generating the language model of the target language,
and for generating a good word alignment model, then MOSES decoder can be used easily for
translation.
For the current work, the MOSES system is trained and a language model is developed using
parallel corpus UMC002. The corpus is published by Charles University in Prague, UFAL. It
contains nearly 45000 parallel sentences. Apart from that, there are parallel tokens for both the
languages. These tokens are obtained from Wikipedia named entries for the years 2008 and 2009.

The translation of individual phrases using MOSES is as follows:

Table 5 Phrase translation with MOSES

66
 International Journal on Natural Language Computing (IJNLC) Vol. 6, No.3, June 2017
Google translator

In the absence of a good training corpus, Google translator /any other statistical translator can be
used to translate the individual phrases. The translation of individual phrases using Google is as
follows:
Table 6 Phrase translation with Google

It has been observed that the negative verb phrase is not translated by Google translator properly.
For example consider few negative verb phrases and the corresponding translations by Google
translator
Table 7 Negative Phrase translation by google translator

Note that the main verb is not at all translated in English to go in the above phrases. Because
of this drawback of the translators, for translating the negated verb phrases, the main verb is
individually translated and appended with a negation. For example for translating

While appending the negation (not) before the translated main verb, the grammatical rules for
English language needs to be considered. The current scheme only considers the above rule for
generating the translation of negated verb phrases. The scheme can be extended further to cater
other verb types.

e) Rule based Phrase Recombination

The recombination of the individual translated phrases must be done based on the ordering rules
derived after studying a variety of Hindi sentences, and the corresponding English translations
from the UMC parallel corpus. The sentences considered for generating the recombination rules
are based on Hindi sentences having different noun case endings and in different order. Some of
them are as follows:

67
 International Journal on Natural Language Computing (IJNLC) Vol. 6, No.3, June 2017
Example:

Interrogative sentences

Negative

The following are the recombination rules that were generated for the various types of sentences.

Table 8 Recombination rules for Simple Positive sentences

Hindi Sentence Structure based on Noun Case Corresponding Translated English Sentence
Ending Structure
Kartaa(doer) ne Karan(instrument) se Kartaa(doer) VERB karma (locus of the result of
Sampradaan(beneficiary) ke liye karma (locus of the action) from Karan(instrument) for
the result of the action) ko VERB Sampradaan(beneficiary)
Kartaa(doer) ko Karan(instrument) se karma (locus Kartaa(doer) VERB karma (locus of the result of
of the result of the action) VERB the action) to Karan(instrument)
Kartaa(doer) ne ((sambandh(relation) kii) Kartaa(doer) VERB karma (locus of the result of
adhikaran(location)) mein karma (locus of the result the action) in adhikaran(location)+
of the action) VERB (sambandh(relation)

68
 International Journal on Natural Language Computing (IJNLC) Vol. 6, No.3, June 2017
Kartaa(doer) (sambandh(relation)) Kartaa(doer) VERB karma (locus of the result of
apaadaan(source) se karma (locus of the result of the action) from apaadaan(source) +
the action) VERB (sambandh(relation))
Kartaa(doer) ne Sampradaan(beneficiary) ko Kartaa(doer) VERB Karan(instrument) to
Karan(instrument) VERB Sampradaan(beneficiary)
Kartaa(doer) ADVERB karma (locus of the result of Kartaa(doer) ADVERB+VERB karma (locus of
the action) VERB the result of the action)
Kartaa(doer) ((sambandh(relation) ke Kartaa(doer) VERB at adhikaran(location) +
adhikaran(location)) par VERB sambandh(relation)

Table 9 Recombination rules for Interrogative sentences

Hindi Noun Case Ending Corresponding English Sentence Structure

Kartaa(doer) (kyu/kyaa/kahaa/..) karma (locus of the Why/What/Where is Kartaa(doer) verb on karma
result of the action) se VERB (locus of the result of the action)
Kya Kartaa(doer) karma (locus of the result of the Do Kartaa(doer) VERB karma (locus of the
action)ko VERB result of the action)
sambandh(relation)Kartaa(doer) kyaa VERB What VERB sambandh(relation)+ Kartaa(doer)

Table 10 Recombination rules for Negative sentences

Hindi Noun Case Ending Corresponding English Sentence Structure

Kartaa(doer) karma (locus of the result of the Kartaa(doer) NEG VERB karma (locus of the
action) NEG VERB result of the action)
Kartaa(doer) sambandh(relation) kaa/ke/kii Kartaa(doer) NEG VERB to
karma (locus of the result of the action) ko sambandh(relation) + karma (locus of the result
NEG VERB of the action)

The phrases are to be recombined using filler words like for, in etc. The filler words are decided
based on the Noun phrases case ending/karaka type. There can be more than one filler word for
the same karaka also. For example at, on, in are the common filler words for the adhikaran
(location) case ending. For example:
Ram is sitting at his friends place.
Ram is sitting on the table.
Ram is sitting in the room.

This results in more than one translation of each sentence, based on the filler word used in each of
the translations. Based on the English language model obtained from SRILM, the top 5
translation are picked from the output list.

As per the recombination order described, and after choosing the best translation based on the
English language model obtained from SRILM, the translation for the example Hindi sentence

using MOSES and Google is as follows:

MOSES: Earth does suns one round in every 12 months.
Google: Earth applies a round of sun in every 12 months.

For interrogative sentence containing kyaa (what), kyuu (why), kisliye (why) the recombination
rule, 9a will be applicable. For interrogative sentence containing kaun (who), recombination rule
9b will be applicable. For example, consider the following sentence:
69
 International Journal on Natural Language Computing (IJNLC) Vol. 6, No.3, June 2017

The phrases identified are

The translation of each phrase, when recombined using the recombination rule of 9a (as kyuu/why
rule is to be considered) results in the following translation

MOSES : Why is Ram going to Sitas house from his house.

Google : Why is Ram going to house of Sita from own house.

The scheme can be further extended if other rules are found to be suited for the phrases and
thereafter generate the best translation based on the language model of the target language.

4. EXPERIMENTAL RESULTS

Experiments are conducted on a set of 5000 sentences from the UMC parallel corpus of nearly
45000 parallel sentences. From the corpus of 45000 sentences, 40000 sentences were analysed to
create the recombination rules explained in Section 3, and remaining 5000 sentences were used
for translation evaluation purpose. The individual phrases are translated using two translators:
MOSES decoder and Google translator.

Following are the outputs of a few sentences using the proposed translation scheme, and using
MOSES and Google, for translation the individual phrases.

70
 International Journal on Natural Language Computing (IJNLC) Vol. 6, No.3, June 2017
The BLEU score is used for the evaluation of the results. The Table 11 depicts the average BLEU
score of the top 5 translations produced for 5000 sentences by the proposed scheme using the two
translators for the phrases, MOSES and Google, and the corresponding comparison with the state
of the art MOSES translator, Google translator and the translation proposed by [12]. The average
BLEU score of the best translation using the proposed scheme with MOSES is 26.5 and with
Google is 27.9. The proposed scheme is producing top-5 output options (which are having BLEU
score of ~26.5 to 14.6 using MOSES and 27.9 to 18.5 using Google). The user can see all the
options available and thereafter choose the best as per his judgement.

Table 11 BLEU score for 1000 test sentences

Top n-th Proposed Proposed Scheme MOSES Google IIT Bombay
translation Scheme Using Using Google translator [12]
MOSES
1 26.5 27.9
2 25.4 25.6
3 22.3 23.2 23.3 24.5 25.7
4 19.7 19.2
5 14.6 18.5

The average BLEU score of the best translation obtained from the proposed scheme (26.5 and
27.9) outperforms when compared with classical MOSES (23.3), Google(24.5), and the IIT
Bombay translator (25.7). The reason can be that, in the current work the relative ordering of the
translated phrases is done using the recombination rules proposed in Table 8,9 and 10. These
rules preserve the semantic information of the sentence and hence the quality of translation
improves; which is not the case for the other translators.

5. CONCLUSION AND FUTURE WORK

The scheme provides a starting point for a Hybrid approach for translating Hindi sentences to
English, with an attempt to preserve the semantic information of the source language sentence.
For the current work the sentences in the problem domain are simple Hindi declarative sentences,
negative sentences, and Content-question interrogative sentences. Since the semantic association
between the syntactic phrases is the basis of the work, the scheme works well for these sentences.
The scheme must be verified for complex, compound and exclamatory sentences. Since these
sentences have different structure, the recombination rules will vary depending on the input
sentence structure.
In the proposed scheme, the filler words (e.g. for, from, in) between the noun case endings
translations are pre-defined as given in the recombination rule defined in Section 3. Consideration
for other filler words which are applicable for each noun case ending can be done. Also, there can
be other considerations, apart from noun case-endings (karakas) when dealing with sentences
other than simple declarative sentences.
The scheme relies heavily on the quality of the chunker for identifying the Hindi phrases. As a
future study, there can be a comparison between the translation qualities of the output for other
chunkers apart for the Shallow Parser. The Indian languages which are like Hindi in structure can
also utilize from this approach.

71
 International Journal on Natural Language Computing (IJNLC) Vol. 6, No.3, June 2017

REFERENCES

[1] H. Almaghout, J.Jiang, and A.Way. CCG augmented hierarchical phrase-based machine translation In
Proceedings of the 7th International Workshop on Spoken Language Translation, Paris, France,
pp.211--218. 2010.
[2] P. Balajapally, P. Pydimarri, M. Ganapathiraju, N. Balakrishnan and R. Reedy, 2006. Multilingual
book reader: Transliteration, word-to-word translation and full-text translation Proceeding of the
13th Biennial Conference and Exhibition Conference of Victorian Association for Library
Automation Melbourne, Australia, pp: 1-12.
[3] S. Bandyopadhyay, (2004) "ANUBAAD - The Translator from English to Indian Languages", in
proceedings of the VIIth State Science and Technology Congress. Calcutta. India. pp. 43-51.
[4] Akshar Bharati, Dipti Misra Sharma, Samar Husain, Lakshmi Bai, Rafiya Begam, Rajeev Sangal,
AnnCorra : TreeBanks for Indian Languages Guidelines for Annotating Hindi TreeBank, Language
Technologies Research Center IIIT, Hyderabad, India (version 2.5) 17/09/2012
[5] Akshar Bharti, Vineet Chaitanya, Amba P. Kulkarni & Rajiv Sangal, (1997) ANUSAARAKA:
Machine Translation in stages, Vivek, a quarterly in Artificial Intelligence, Vol. 10, No. 3, NCST
Mumbai, pp. 22-25
[6] Akshar Bharati, R. Moona, P. Reddy, B. Sankar, D.M. Sharma & R. Sangal, (2003) Machine
Translation: The Shakti Approach, Pre-Conference Tutorial, ICON-2003
[7] A. Birch, M. Osborne, and P. Koehn.CCG supertags in factored statistical machine translation. In
proceedings of the Second Workshop on Statistical Machine Translation, June 23, 2007, Prague,
Czech Republic; pp.9-16, ACL 2007.
[8] N. Chatterjee, A. Johnson and M. Krishna. Some Improvements over the BLEU Metric for
Measuring Translation Quality for Hindi. Proc. ICCTA, IEEE Computer Society, 2007, pp. 485
490
[9] S. Dave, J. Parikh, and P. Bhattacharyya. Interlingua Based English Hindi Machine Translation and
Language Divergence. Journal of Machine Translation, 17, September 2002.
[10] C. David. A hierarchical phrase-based model for statistical machine translation. In Proceedings of the
43rd Annual Meeting of the ACL, pages 263270, Ann Arbor, MI, 2005.
[11] V. Dayal. Locality in WH Qantification: Questions and Relative Clause in Hindi. Kluwer Academic
Publishers.1996.
[12] P. Dungarwal, R. Chatterjee, A. Mishra, A. Kunchukuttan, R. Shah, and P. Bhattacharyya. 2014. The
iit bombay hind- english translation system at wmt 2014. ACL 2014, page 90.
[13] Vishal Goyal, Gurpreet Singh Lehal, 2011, Hindi to Punjabi Machine Translation System
Proceedings of the ACL-HLT 2011 System Demonstrations, pages 16, Portland, Oregon, USA, 21
June 2011
[14] D. Gupta and N. Chatterjee Divergence in English to Hindi Translation: Some Studies. International
Journal of Translation, Vol 15. No. 2, pp 5 24, 2003.
[15] R. Jain, R.M.K. Sinha, and A. Jain. Anubharti: Using hybrid example based approach for machine
translation. In Symposium on Translation Support Systems, SYSTRANS, Kanpur, India, February,
2001
[16] G. S. Josan and G. S. Lehal, 2008, A Punjabi to Hindi Machine Translation System Coling 2008:
Companion volume: Posters and Demonstrations, Manchester, UK, pp. 157-160
[17] P. Koehn. Pharaoh: A beam search decoder for phrase-based statistical machine translation models. In
R.E. Frederking and K. Taylor, editors, Proceedings of the American Machine Translation
Association, volume 3265 of Lecture Notes in Computer Science, pages 115124. Springer, 2004.
[18] G. Labaka, C. Espa na-Bonet, L. Mrquez, K. Sarasola: A hybrid machine translation architecture
guided by Syntax. In Machine Translation, 28 (2014), pp. 135
[19] Ananthakrishnan R, Kavitha M, Jayprasad J Hegde, Chandra Shekhar, Ritesh Shah, Sawani Bade &
Sasikumar M., (2006) MaTra: A Practical Approach to Fully- Automatic Indicative English Hindi
Machine Translation, In the proceedings of MSPIL-06.
72
 International Journal on Natural Language Computing (IJNLC) Vol. 6, No.3, June 2017
[20] R. Sinha, and A. Jain, Angla Hindi: An English to Hindi Machine-Aided Translation System, MT
Summit IX, New Orleans, USA, 23-27 Sept. 2003
[21] R. Sinha and A Thakur. On Translation of Interrogative sentences from Hindi to English,
International Conference on Machine Learning; Models, Technologies and Applications, MLMTA,
Las Vegas, Nevada, USA, 2006.
[22] R. Mahesh K. Sinha & Anil Thakur, (2005) Machine Translation of Bi-lingual Hindi-English
(Hinglish) Text, in proceedings of 10th Machine Translation Summit organized by Asia-Pacific
Association for Machine Translation (AAMT), Phuket, Thailand
[23] G. Singh and D. Lobiyal. A Computational Grammar For Hindi Verb Phrase, Proc. International
Conference on Expert Systems for Development, 1994, 28-31 Mar 1994, pp: 244-249, DOI:
10.1109/ICESD.1994.302273.
[24] H. Somers. Review article: Example-based machine translation. Machine Translation, 14(2):113157,
1999. McGregor, R. S. 1995. Outline of Hindi Grammar [3rd edition]. Oxford: Oxford University
Press.
[25] H. Uchida and M. Zhu. Interlingua for Multilingual Machine Translation. In Proceedings of the
Machine Translation Summit IV, Kobe, Japan, July 20-22, 1993.
[26] P. Williams, R. Sennrich, M. Nadejde, M. Huck, E. Hasler, and P. Koehn, Edinburghs Syntax-
Based Systems at WMT 2014, in Proc. of the Workshop on Statistical Machine Translation (WMT),
Baltimore, MD, USA, June 2014, pp. 207214.
[27] A. Zollmann, A. Venugopal, S. Vogel and A. Waibel. The CMU-UKA Syntax Augmented Machine
Translation System for IWSLT-06. In Proc. of International Workshop on Spoken Language
Translation (IWSLT-06) -- system papers, Kyoto, Japan. 2006.
[28] https://en.wikipedia.org/wiki/Evaluation_of_machine_translation#Automatic_evaluation
[29] A. Kalyani, H. Kamud, S. Pal Singh, and A. Kumar. Assessing the quality of mt systems for hindi to
english translation. In International Journal of Computer Applications, volume 89, 2014.

AUTHORS
Susmita Gupta is a doctorate student in the in the Department of Mathematics, IIT
Delhi. Her primary research areas include Machine translation, typically Example Based
Machine Translation, and Hybrid approaches for translation, primarily from Indian
languages to English. She had done her M.Tech in Computer Application from IIT Delhi,
and B.Tech in Computer Science from Institute of Engineering and Technology,
Lucknow.

Dr. Niladri Chatterjee is a Professor of Statistics and Computer Science in the

Department of Mathematics, IIT Delhi. His primary research areas are:Natural Language
Processing, Semantic Web, Statistical Modeling, His association with IIT Delhi is
closed to 15 years. Prior to that he had worked in the Dept. of Computer science,
University College London, and at Indian Statistical Institute, Calcutta. In 2010 he has
been a Visiting Scientist in Dipartimento di Informatica, University of Pisa, Italy. He has
over 70 publications international and national journals and conferences. He has been the Organizing Chair
of CICLING 2012 in March 2012.

73