Professional Documents
Culture Documents
Machine Translation and Its Approaches: Vanlalmuansangi Khenglawt, Lal Anpuia
Machine Translation and Its Approaches: Vanlalmuansangi Khenglawt, Lal Anpuia
Language is a medium of communication for different cultures around the world. The communication
between two cultures is blocked by language barrier. To solve the problem machine translation approach is
widely used which is a software used to translate a text or speech from one language to another. It can be
referred to as automatic or instant translation as it is much faster in comparison to human translation.
Machine translation plays a very important role and in highly applicable in both individual and business
areas. The translation system can be either bilingual or multilingual. Most of the machine translation is
bidirectional translation which can be either translated from source language to target language or from
target language to source language. Otherwise it is unidirectional which can be translated only from source
language to target language. It has many approaches for automatic translation of language. This paper
focusses on three different approaches, the rule-based machine translation, the corpus-based machine
translation and the hybrid-based approach. Each of these approaches has its own pros and cons. We also
present several challenges faced in the translation of languages. Also, the metrics for measuring the
performance or the accuracy of the machine translation output have been mentioned. There are still several
works required to develop completely automatic translation system. The performances of the translation
system are still unsatisfactory especially in low resource language. As language is evolutionary in nature,
there are still vase areas to cover in machine translation system.
Keywords: Machine translation, language, rule-based machine translation, corpus-based machine translation,
hybrid-based translation.
CHALLENGES OF MACHINE TRANSLA- linguistics knowledge to the language. However not only
it requires grammar, linguistics and vocabulary it also
TION requires the knowledge gathered from the past experi-
enced.
Translation of a language is not a simple task. There
are several challenges to be dealt with when translating Table 4: Pronoun resolution.
one language to another. The various challenges are as
Sentence Pronoun (it)
follows:
resolution
The computer outputs the data, it is fast It refers to
1. Lexical ambiguity the computer
The computer outputs the data, it is It refers to
Lexical ambiguity is one of the main challenges in stored in ascii the data
translating one language to another. The words in the
source language can have more than one meaning as
shown in Table 1. Also a group of words or a complete 4. Approaches for machine translation
sentence can have more than one meaning as shown in
Table 2. For a machine to understand and translate accu- A machine translation system is broadly classified into
rately the lexical ambiguity needs to be resolved and it is three approaches. Rule based machine translation; Cor-
a great challenge to translate the language correctly. pus based machine translation and Hybrid machine
translation. The Rule based machine translation is further
Table 1: Word with different meaning. classified into direct translation, transfer based transla-
tion and Interlingua translation. The Corpus based ma-
English Mizo chine translation is also further classified into statistical
Read a Book Lehkhabu chhiar machine translation and example based machine transla-
Book the flight ticket Thlawhna ticket hauh tion. The Hybrid machine translation is also further classi-
rawh
fied into ruled based machine translation guided hybrid
and statistical machine translation guided hybrid. The
Table 2: Sentence with more than one meaning.
different approaches of machine translation have been
Sentence Meaning summarized in Figure 1.
I saw bats 1.Bats are animals which can fly Translation of one language to another by a machine
2. Multiple cricket bats translation requires analysis of the source language and
generates an output or target language by using anyone
of the machine translation approaches.
2. Differing word order
RULE-BASED MACHINE TRANSLATION
Two languages may have different word order or dif-
ferent structures of word. For example, English language
The rule-based machine translation is also known as
used a SVO structure (subject-verb-object) whereas Mizo
knowledge driven approach. This approach is the first
language used OSV structure (object-subject-verb) as
approach developed in the field of machine translation
shown in Table 3.
and it is based on linguistic information. The translation
Table 3: Differing word order. system consists of a collection of grammar rules, a lexi-
con and software programs to process the rules (Antony,
Language Sentence Structure 2013; Benson et al., 2013). It produces more predictable
English I eat rice SVO output for grammar since it deals with syntactic, seman-
Mizo Chaw ka ei OSV tic and morphological analysis in both source language
and target language. Building ruled based machine trans-
lation is expensive as all the rules of the language need
3. Pronoun resolution to be applied and there is a requirement of huge linguis-
tic knowledge. But once it is built it can be deeply ana-
The pronoun is a substitution of a noun or noun lysed at syntax and semantic level.
phrase. It can refer to either the subject or the object of There are three different types of ruled based ma-
the sentence as shown in Table 4. So depending upon chine translation. They are direct translation, transfer
the sentence used the pronoun needs to be resolved based translation and Interlingua translation. The levels
which is very challenging task for machine translation. of analysis of the three types are shown through vau-
quois triangle in Figure 2 (Dorr et al., 2004; Benson et al.,
To build a translation machine it requires having good 2016).
142
Advances in Engineering Research, volume 178
3. Interlingua translation
143
Advances in Engineering Research, volume 178
tical machine translation consists of three models known as statistical smoothing and automatic post edit-
(Benson et al., 2016): language model, translation model ing.
and decoder model. The language model gives possible
translation for each word or phrase in the input sentence 2. Statistical machine translation-guided hybrid
with a probability assigned to each translation P(T). The
translation model compute the conditional probability of Rules are adapted in the corpus at the pre-processing
target sentences by giving the source sentence P(T/S). stage to guide the statistical engine. Rules are also used
The decoder model search for the best translation possi- at the post processing stage or at the core model of the
ble P(S,T) by maximizing the product of two probabilities, system to normalize the performance (Costa-Jussa et al.,
the language model and translation model as in the 2016). These approaches avoid the needs of creating set
equation: of linguistic rules by extracting those rules from the train-
ing corpus. The drawback is still the same as normal sta-
P(S,T)=argmax P(t)*P(T/S) tistical machine translation that the accuracy of the
translation depends solely of the similarity of the input
2. Example-based machine translation test to the text of the training corpus. (Sandeep and
Chang, 2015).
The translation system uses the corpora to find analo-
gous examples between the source language and the PERFORMANCE METRICS FOR EVALUA-
target language. It is also called memory based transla-
tion. It uses a point to point mapping with a similarity TION
measures such as word, syntactic or semantic similarity
to identify the approximately matching sentence. The There are several evaluation methods which have
translation system is categorised into two modules: re- been used for measuring the performance of the ma-
trieval module and adaption module (Sandeep, 2015). chine translation output. The evaluation of machine
For a given input sentences, the retrieval module re- translation contains both manual and automatic evalua-
trieves the similar sentences and it translation from the tion methods. Some of the automatic evaluation metrics
corpus. From the retrieval module the adaption module are as follows:
finds out the part of translation that can be reused. If the
input sentence and the retrieval sentence match, then Word error rate (WER)
the correct translated output is given. If it does not
match exactly, the relevant match correspond to the It works at the word level. WER can be obtained by
source language is used instead. using editing distance between both the sentences
(Vidal, 1997). WER is the percentage of word that is to be
changed in the translation to produce the desired sen-
HYBRID MACHINE TRANSLATION tence. However, the drawback is the dependency of the
reference sentence. There can be multiple correct trans-
The hybrid approach uses a combination of rule lation of sentence but the metric considers only the ref-
based approach and statistical based approach (Benson erence sentence to be correct.
et al., 2016). Rule based approach has a high accuracy as
it deeply analyse at syntax and semantic level but it is
Sentence Error Rate (SER)
very expensive as it requires a huge linguistic rules. On
the other hand, the statistical based approach is less ex- It is the error metric of machine translation that
pensive as it uses the mathematical reasoning problem measures the number of changes required to match the
but needs huge corpora which are a not available for low reference sentence.
resourced languages. So, in a Hybrid machine translation
the drawback of both the approaches were excluded to
BLEU Score
give a high efficiency (Sanjay, 2010). They are further
classified into two approaches: Rule based machine
BLEU stands for bilingual evaluation understudy. It is
translation guided hybrid and Statistical machine guided
one of the first metrics to claim a high correlation with
hybrid.
human judgement of quality. The BLEU score is calcu-
lated in terms of sentences. It measures how many word
1. Rule-based machine translation-guided hybrid sequences in the sentence under evaluation match the
word sequences of some reference sentence. It also
The translation is preform using rule based engine. A deals with the penalty for translation with sentences
corpus is introduced to reduce expensive development having a significantly high differs in length comparing to
of linguistics rules. Statistical models are used to correct the reference translation.
or adjust the output from the rules engine. It is also
144
Advances in Engineering Research, volume 178
The METEOR metric Chang, J.S., Su, K.S. (1997). Corpus based statistics ori-
ented machine translation researches in Taiwan.
The metric is based on the harmonic mean of uni- AMTA. 165-173.
gram precision and recall, with recall weighted higher Cheragui, M. A. (2012). Theoretical overview of machine
than precision. It also has several features like stemming translation. Proceedings ICWIT.
and synonymy which are not present in other metrics. It Costa-Jussa, M.R., Jose, A.R., Fonollosa (2015). Latest
is design to produce a good correlation at the sentence trends in hybrid machine translation and its applica-
or segment level tions. Computer Speech and Language 32(1): 3-10.
Dorr, B.J., Eduard, H., Lori, L. (2004). Machine translation:
CONCLUSION interlingual methods. In: Encyclopedia of Language
and Linguistics 2nd edition, p. 939.
In this paper we have reviewed various machine Harjinder, K., Vijay, L. (2013). A survey of machine trans-
translation approaches, the challenges faced in transla- lation approaches. International Journal of Science,
tion system and the performance metrics for evaluation. Engineering and Technology Research (IJSETR) 2 Issue
The survey shows that there are several approaches to 3.
cater the different needs of translation. However, with Mohamed, H.Z., Shafeen, N. (2017). A brief study of chal-
different approaches the translation performance of a lenges in machine translation. International Journal of
machine is quite low in terms of fluency, fidelity, post Computer Science Issues, 14 issue 2.
edit and precision in comparison to human translation. Sandeep, S., Vineet, S. (2015). A survey of machine trans-
Therefore, we conclude that there are still vast areas to lation techniques and systems for indian languages.
improve the system to utilise a fully automatic transla- IEEE International Conference on Computational Intel-
tion without the need for human evaluation. ligence and Communication Technology.
Sanjay, K. D., Pramod, P. S. (2010). Machine translation
system in Indian perspectives. Journal of Computer
REFERENCES Science. 6, 1111-1116.
Sindhu, D.V., Sagar, B.M., Rajashekar, M. (2014). Survey
Antony, P.J. (2013). Machine translation approaches and on machine translation and its approaches. Interna-
survey for Indian languages. International Journal of tional Journal of Advanced Research in Computer and
computational Linguistics and Chinese Language Communication Engineering. 3, Issue 6.
Processing. 18(1): 47-78. Tripathi, S., Sarkhel, K. (2010). Approaches to machine
Benson, K., Lawrence, M., Wanjiku, N. (2016). A review translation. Annals of Library and information studies
on machine translation approaches. Indonesian Jour- 57: 388-393.
nal of Electrical Engineering and Computer Science 1: Vidal, E. (1997). Finite state speech to speech translation.
182-190. In: Proceedings of the International Conference on
Acoustics, Speech and Signal Processing, Munich, Ger-
many.
145