English-Arabic Hybrid Machine Translation System using EBMT and Translation Memory

Abstract

—The availability of a machine translation to
translate from English-to-Arabic with high accuracy is not
available because of the difficult morphology of the Arabic
Language. A hybrid machine translation system between
Example Based machine translation technique and Translation
memory was introduced in this paper. Two datasets have been
used in the experiments that were constructed by using internal
medicine publications and Worldwide Arabic Medical
Translation Guide Common Medical Terms sorted by Arabic. To
examine the accuracy of the system constructed four experiments
were made using Example Based Machine Translation system in
the first, Google Translate in the second and Example Based with
Google translate in the third and the fourth is the system
proposed using Example Based with Translation memory. The
system constructed achieved 77.17 score for the first dataset and
63.85 score for the second which were the highest score using
BLEU score.

Full text

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 10, No. 1, 2019
195 | P a g e
www.ijacsa.thesai.org
English-Arabic Hybrid Machine Translation System
using EBMT and Translation Memory
Rana Ehab1, Mahmoud Gadallah3
Computer Science Department, Modern Academy for
Computer Science and Management Technology
Cairo, Egypt
Eslam Amer2
Computer Science Department
Misr International University
Cairo, Egypt
Abstract—The availability of a machine translation to
translate from English-to-Arabic with high accuracy is not
available because of the difficult morphology of the Arabic
Language. A hybrid machine translation system between
Example Based machine translation technique and Translation
memory was introduced in this paper. Two datasets have been
used in the experiments that were constructed by using internal
medicine publications and Worldwide Arabic Medical
Translation Guide Common Medical Terms sorted by Arabic. To
examine the accuracy of the system constructed four experiments
were made using Example Based Machine Translation system in
the first, Google Translate in the second and Example Based with
Google translate in the third and the fourth is the system
proposed using Example Based with Translation memory. The
system constructed achieved 77.17 score for the first dataset and
63.85 score for the second which were the highest score using
BLEU score.
Keywords—Hybrid machine translation system; translation
memory; internal medicine publications; google translate; BLEU
I. INTRODUCTION
In 1952 the first conference on MT came. There was the
first demonstration of a translation system in January 1954,
and it attracted a great deal of attention and since then there
has been no stopping [1]. Since Language technologies are
very successful nowadays Machine Translation has been
applied to the medical domain [2]. The quality of language
technologies is growing very rapidly [2]. People with different
languages can share ideas and information worldwide on
every topic as business, economic, educational, political,
socio-cultural, etc. if machine translation researchers have the
ability to develop a perfect multilingual machine translation
system [3]. The presence of a machine translation that has the
ability to translate any text in any domain at the required
quality is expected in not-too-distant future [2]. Machine
translation must present a reasonable approach to translate
terms to meet commercial needs [4]. Generally users are

it means [2]. However, some applications require much more
than this [2]. As example, in the medical field the beauty and
correctness of the text may not be important, but the precision
and efficiency of the translated message are very important
[2]. Machine translation systems can be used to translate
medical records [2].
The most important task for saving with high-quality
medical services is the communication between medical
physicians and patients [5]. If medical physicians and patients
do not share a common language, the diagnosis and treatment
will be more difficult due to the language barrier that prevents
effective communication [5]. Another case is people who
travel to receive high-quality or affordable medical treatment
that is not available in their home country [5]. When
translating medical information and make it understandable
both physicians and patients will benefit [6]. As an example,
Healthcare Technologies for the World Traveller confirm that
a foreign patient may need a description of their diagnosis
with a related and full set of information [2].
The world has become a small village because of the rapid
changes in information and communication technology via
internet where people from all over the world can connect
with each other in dialogue and communication [7]. The
translation databases and translator workstation such as the
Google Translate (GT), Bing Translate, Yahoo, Babel Fish
and Systran that were developed and influenced by the
internet was the development of computer-based translation
tools [8].
Using websites in translation has been outspread.
However, the task of translating a medical text is not as easy
as translating any other English text because of the complex
information that it contains. So, using existed systems in
translation a medical text produces a translation text with
some problems. Because of the difference in language
categories, current methods are far from being at the degree
where they can be of practical use especially in English-to-
Arabic medical translation.
In the medical domain most institutional and research
information is available as English text [9]   
know English language well they will not be able to make use
of these information without a help [10]. So, the task is
helping everyone to use web and this will be achieved by
automatic language translators [10]. Because of the flow of
information in foreign languages through web the use of
machine translation technology is must [11].
Most of the researches in Arabic Machine Translation are
mainly concentrated on the translation between English and
Arabic because English is a universal language [12]. This will
help in simplifying the Arab communication with other
countries [12]. That was the reason to choose translating from
English to Arabic.

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 10, No. 1, 2019
196 | P a g e
www.ijacsa.thesai.org
The field of Machine Translation research is largely
controlled by corpus-based nowadays, or data- driven
approach [13]. Although Example Based Machine Translation
(EBMT) and Statistical Machine Translation (SMT) are from
corpus-based model each of them has their own advantages
and disadvantages [14]. Example Based Machine Translation
can works well with a limited training and testing datasets

other than Statistical Machine Translation that needs a large
dataset to result a significant translation [15]. Also when the
nature of the training and test are close the Example Based
Machine Translation System works well. Also reusing the
segment of a test sentence that can be found in the source side
of the example-base improves the translation by Example
Based Machine Translation Systems. The idea of Example
Based Machine Translation is getting translation examples of
similar sentences.
Using Example Based Machine translation is often linked
      
[15]. Translation memory (TM), is a database that is today
widely used Computer-Assisted Translation (CAT) tool
prepared for future reuse of already translated texts [16]. The
similarity between them is they both reuse the examples from
the existing translations. The main difference between them is
that Example Based Machine Translation is an automated
technique for translation whereas Translation Memory is an
interactive tool for the human translator [15].
Beside the technique used to build the translation system
the dataset that is used in training and testing of the system is
important. Some machine translation systems evaluation was
low because of the dataset used. So, when building a machine
translation system it is very important to consider the
goodness of the dataset that will be used. So, in the
experiments in this paper two datasets were used. The first
dataset constructed using internal medicine publications from
[17]. The second one constructed using internal medicine
publications and Worldwide Arabic Medical Translation
Guide Common Medical Terms sorted by Arabic which is an
English-Arabic medical dictionary that will be described later.
The attempt to use Example based machine technique and
Translation memory to translate English medical text to
Arabic medical text will be described in this paper. As the
constructed datasets were not large the choice to use Example
based machine technique that works well with a limited
training and testing datasets was the best. Also with the
advantages of Translation memory which are consistency,
speed and cost-saving [16] that will benefit the resulted
translation. Also, from the benefits of using Translation
memory: need for consistent use of terminology, data sharing
of common resources, re-use of already translated and revised
text suggest used of Translation Memory, in its simplest form
a database [16].
The rest of the paper is organized as following : the second
section describes the recently related works in machine
translation in medical domain and non-medical domain, the
third section describes the issues that face medical domain,
the forth section describes the datasets and the hybrid
translation system from English to Arabic that was built, the
fifth section describes the experiments to evaluate the system
built, the sixth section shows the evaluation of the experiments
using BLEU metric and finally the last section shows a brief
conclusion of the work.
II. RELATED WORK
In the medical domain there are many machine translation
systems for various languages have been developed using
different approaches of machine translation. Also machine
translation has been developed to translate English text to
Arabic text but not in medical domain.
Dandapat, et al. [15] used Example-based machine
translation and Translation Memory to translate medical text
from English to Bangle. They translated receptionist dialogues
of medical and primarily appointment scheduling. Their first
step was to collect their data and then building a Translation
Memory automatically from a corpus of patient dialogue using
Moses toolkit. They created two Translation Memories the
first contains phrase pairs that are aligned and the second one
contains the word aligned file [15].
They made five different experiments to show the
accuracy of their system. The fourth experiment achieved the
highest accuracy which is 57.56 [15] where they used their
system with the first and second Translation Memories.
However they achieved the highest accuracy, some errors
appeared, the first was the wrong of source-target equivalent
in both Translation Memory systems [15]. The second in the
recombination step that some words are translated separately
[15].
Névéol, et al. [18] built a statistical machine translation
system to translate systematic reviews from English to French.
They used three different datasets. They made five systems.
During the evaluation the last system achieved the best
accuracy which was (40.00 BLEU) [18] where they used
Cochrane translation table and an integrated translation table
between EMEA and WMT. Also, Subalalitha, et al. [19] tried
to use statistical machine translation to translate from English
to Hindi and achieved accuracy (73.43).
Renato, et al. [20] discussed translating clinical term
descriptions from Spanish to Brazilian Portuguese. HIBA
dictionary was used as a Spanish dictionary. They collected
medical terms of Portuguese language using several sources.
They made two experiments and evaluated them. For both
experiments they used for translation Bing, Google Translate
and their system M-SMT. In the first experiment their system
achieved the highest score which is (58.9) [20] using BLEU
score. In the second experiment their system achieved (86.7)
[20]. That shows that their system achieved the highest score.
As showed that the second experiment achieved higher
scores in all translation systems. However although they
achieved high scores there were some errors as [20]: OOV
words are usually translated into English or left in Spanish, a
part of the corpus had words with spelling in European
Portuguese, Compound medical terms, especially drugs with a
hyphen, possibly misaligned in training.
Li, et al. [21] developed a hybrid translation system
between Dictionary based machine translation technique and
Statistical machine translation technique. They translated

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 10, No. 1, 2019
197 | P a g e
www.ijacsa.thesai.org
query terms in medical domain from English to German and
vice versa [21]. Their corpus was a mix from more than one
corpus. They made two experiments to evaluate their system.
In the first one they used Phrase-based machine translation
system and in the second one they used their system [21].
According to their evaluation they achieved better evaluation
than the first one. Their system achieved (15.3) for translating
from English to German and (24.5) from German to English
which was higher than the accuracy of the other system.
, et al. [22] goal was to translate medical data from
English to Polish and vice versa so they developed a SMT
system for this purpose. For their dataset they used the
European Medical Agency (EMEA) data. To evaluate their
system they made 13 experiments. The results showed that
translating from Polish to English evaluates better than
translating from English to Polish [22] The fifth experiment
achieved the highest score among the other experiments that
was (76.34) for BLEU score for translation from Polish to
English and (73.32) for translating from English to Polish
[22]. Also Johanna Johnsi Rani G, et al. [23] used SMT to
translate medical reports from English to Tamil. They
evaluated their system with the results of Google Translate
and they achieved better accuracy.
, et al. [2] built a machine translation system rely on
using neural network. They used European Medicines Agency
(EMEA) parallel corpus to derive their corpus. The system
translates Polish medical text into English medical text and
vice versa [2]. They made three experiments to evaluate their
system. Their system achieved (24.32) for translating from
Polish-to-English and (17.50) for translating from English-to-
Polish [2] that was lower than the other two experiments. Also
Artetxe, et al. [24] tried to use neural network to translate
from French and German to English but they achieved low
accuracy.
Amer, et al. [25] built a query translation system which is
Wiki transpose for cross-lingual information retrieval (CLIR)
that relied on Wikipedia as a source for translations. They
used the system to check how reliable Wikipedia is to get
corresponding translation coverage of English to Portuguese
and also Portuguese to English queries [25]. For their
evaluation they made two experiments. They used English
Open Access, Collaborative Consumer Health Vocabulary
Initiative dataset in the first experiment [25]. They used a
collection of Portuguese medical terms that were rated by
medical experts as medical terms in the second experiment. A
coverage ratio in Wikipedia about 81% and about 80% [25] in
single English and Portuguese terms respectively was reached.
Rana Ehab, et al. [17] built a machine translation system
using Example based machine translation technique to
translate English medical sentences to Arabic medical
sentences. They constructed their parallel corpus using the
internal medicine publications for internal diseases only [17].
The matching stage was used from Example based technique
to find the closest example from the parallel corpus as the
example based for the system. The second experiment made
using Google translate and the same data were translated to
examine the accuracy of their system but Google translate
achieved higher score than their system. Google translate
achieved (53.56) for BLEU score and their system achieved
(48.86) [17].
Shaalan, et al. [12] built a translation system to translate
English noun phrase into Arabic. They used Transfer machine
translation approach as their system [12]. They evaluated their
system by using 50 titles from the computer science domain as
training dataset for their system and for testing they used other
66 new real thesis titles from the computer science domain.
Their evaluation showed that the system translated 47 noun
phrases correctly and the remaining 109 noun phrases have
problems [12].
Shaalan, et al. [26] built a translation system using Rule-
based transfer machine translation technique to translate
expert systems in the agriculture domain from English to
Arabic and vice versa. This translation process includes
translating knowledge base, in particular, prompts, responses,
explanation text, and advices. Those expert systems are built
in CLAES
1
[26].
They used for their system a set of real parallel 100
phrases and sentences from both English and Arabic versions
of agricultural expert systems at CLAES that were used as a
gold standard reference test data [26]. They made the
evaluation through two experiments. The second experiment
achieved higher accuracy than the first which is 0.6427 for
English to Arabic direction and 0.8122 for Arabic to English
direction [26]. Also Kouremenos, et al. [27] used also Rule-
based technique to translate Greek to Greek Sign language.
Al-Taani, et al. [28] translated well-structured English
sentences into well-structured Arabic sentences using rule
based approach. They used 184 English proverbs from Al-
Mawrid, English- Arabic dictionary [28]. Also they used 125
well structures English sentences from many text books.
During the evaluation 57,3% of the first dataset translated
correctly and 84.6 of the second dataset translated correctly
[28]. These results were not as they supposed because of many
reasons. From these reasons that proverbs have no specific
structure, also proverbs are much related to the culture of
some nations [28]. Also Mouiad Alawneh, et al. [3] translated
well-structured English sentences into well-structured Arabic
sentences but using Grammar parser and example based
machine translation technique.
As shown in the previous approaches of machine
translation in medical domain most of them used Statistical
machine translation technique and Example based machine
translation technique. There was also an attempt to use neural
network in translation but in comparison with SMT the second
achieved higher score. Also an approach [15] used Translation
memory with Example based technique and achieved higher
scores than using Example based technique with SMT. For
this reason the proposed system is to build a system using
Example based machine translation and Translation memory.
Also as shown that most of English to Arabic machine
translation systems in non-medical domain used Rule based
machine translation approach as they need to analyze the
English text in terms of morphology, syntax and semantic
1
Stands for Central Laboratory of Agricultural Expert Systems (CLAES),
Agricultural Research Centre (ARC), Egypt, http://www.claes.sci.eg

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 10, No. 1, 2019
198 | P a g e
www.ijacsa.thesai.org
which is not important for English text in medical domain.
The strengths of rationalism method and empiricist method are
merged through using Hybrid machine translation [29].
III. ISSUES WITH MEDICAL DOMAIN
In [17] to construct an efficient Machine Translation
system for a Medical Domain there are two main issues which
are: parallel corpus collection, size and type of corpus. Beside
them there is a third issue which is building a Translation
Memory [15]. The medical terms are different from any other
English terms. For that building an efficient medical corpus is
not an easy task. To evaluate the system two datasets were
used. The first one is [17] where they used internal medicine
publications to build it. The second dataset constructed using
internal medicine publications and Worldwide Arabic Medical
Translation Guide Common Medical Terms sorted by Arabic
which is an English-Arabic medical dictionary to build
English-Arabic parallel corpus. The first corpus consists of
259 medical sentences; for each sentence there are 8 words on
average [17]. The second corpus consists of 509 medical
sentences.
The proposed system uses Example-Based Machine
Translation which is a data-driven machine translation
technique [15, 17] that needs a machine readable parallel
corpus. So when building such a system how many examples
needed musy be known? In a comparison with such systems
the first corpus is very small but the second corpus is larger
than other corpuses as in Tabel 1. The first corpus is small
because it is built from only the medical data of internal
diseases but the second corpus includes more diseases besides
using Worldwide Arabic Medical Translation Guide Common
Medical Terms sorted by Arabic. As seen in Table 1. many
systems have been constructed using a small corpus.
As there is no access to an existed Translation Memory
building a Translation Memory automatically for the proposed
system using
2
Moses toolkit was considered. A Translation
memory was created based on word aligned file created using
Moses word alignment (Giza++) [15]. Because each source
word has multiple target equivalents all the multiple
equivalent words in sorted order were kept. This Translation
memory will help in the second stage of the system which is
finding the alignment between the result from the database
and its translation.
TABLE I. SOURCE OF MEDICAL TERMS OF PORTUGUESE LANGUAGE
System
Language Pair
Size
TTL
English-> Turkish
488
TDMT
English->Japanese
350
EDGAR
German-> English
303
ReVerb
English-> German
214
ReVerb
Irish -> English
120
METLA-1
English -> French
29
METLA
English -> Urdu
7
2
Moses (http://www.statmt.org/moses/) is a SMT system that
automatically trains a translation model for any language pair.
IV. OUR APPROACH
A. Data Preparation
In each domain words have different meanings so, their
translation has to fit in the excepted representation in the
domain. Therefore to ensure that they are treated consistently
throughout the technical text, it is important to identify them
correctly [30].
In the previous section, as mentioned two datasets were
used. The first one was constructed by [17] where they built it
from the indications and side effects from the internal
medicine publications in both languages English and Arabic
for internal diseases only.
The second dataset were constructed from indications and
side effects from the internal medicine publications for
multiple diseases and Worldwide Arabic Medical Translation
Guide Common Medical Terms sorted by Arabic .After that,
some processing on English data were made as tokenization, a
lower casing, and final cleaning. Pre-processing Arabic
sentences could change the meaning of the sentence due to the
morphology of the language and the meaning of the sentence
is very sensitive in the medical domain .So, no pre-processing
for the Arabic will be done.
B. Translation System
In the example based translation, a system is defined
which contains a set of source language sentences and
corresponding target language sentences. During the run time,
example based translation use bilingual corpus as its database.
This database is stored in the translation memory. In
translation memory, the user translates text these translations
are added to a database, and when the same sentence occurs
again during the translation, the previous translation is
inserted in to the translated document. The advantage of the
example based translation the translation memory saves the
user effort of re translating the sentence and this saves the
processor time and also the user time. EBMT can help to
overcome some of the weaknesses of the other approaches
[31].
With the advantages of the Example Based Machine
Translation approach and the Translation Memory a hybrid
system that uses both of them to translate English medical
sentence to Arabic medical sentence was developed. Arabic
language was chosen as destination language because there is
many possible ways to express the same sentence in Arabic
that provides a significant challenge to MT [3]. The accents of
modern Arabic are well-known as having agreement
asymmetries that are sensitive to word order effects. As all
Example Based Machine Translation system the proposed
approach is from three stages which are: Matching, Adaption
and Recombination [15].
1) The proposed hybrid machine translation system: The
hybrid machine translation system in Fig. 1 is used to translate
medical sentences from English to Arabic using Example
Based machine translation and Translation memory.
User initially inserts the input English sentence, the
sentence then goes to some pre-processing steps: tokenization,
lower casing and stop word removing, then the sentence sent

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 10, No. 1, 2019
199 | P a g e
www.ijacsa.thesai.org
to Example based which is the parallel corpus to find the
closest example by computing edit distance between the input
sentence and each example and this will be discussed later.
The example that gets the highest score will be the closest
example. Then using the parallel corpus the translation of the
closest example will be gotten.
Then the alignment between the input sentence and the
example will be found to find the unmatched portions, this
done while computing the edit distance. Also the alignment
between the example and its translation example will be found
to find the unmatched portions by using the Translation
memory and this will be described in section 4.2.3. Then the
unmatched portions of the input sentence will be replaced with
the unmatched portions of the translated sentence and add or
substitute from the translated sentence and this will be
discussed later.
Finally the un-translated segments that were replaced will
be translated and added to the translated sentence using the
translation memory and then the final translated sentence was
get.
2) Matching stage: In this stage the task is to find the
source closest examples from the database that closely
matches the input sentence and that is done by using word-
based edit distance metric (1) (Levenshtein, 1965; Wagner and
Fischer, 1974) [16].
Score (Si,Se)= 1- 
 || (1)
Where Si denotes the input sentence and Se denotes the
example from the database sentence. So, |Si| and |Se| denotes
the length of an input sentence and example sentence
extracted from database and ED(Si,Se) refers to the word
based edit distance between Si and Se.
Based on the above scoring technique the following
examples from the database in (2) for the input sentences in
(1) were gotten.
(1) a- impaired function of the liver
b- arthrosis
c- nasal congestion
(2) a- impaired function of the kidneys
b- arthritis
c- lung congestion
Then the associated translation St in (3) was gotten for the
sentences in (2) from the database. This translation will be
used in the following subsections to get new translation texts.
(3) a-

b-

c-

3) Adaption stage: In this stage the unsuitable fragments
from the resulted translation from the previous stage were
extracted. For this purpose the three sentences that have gotten
from the previous stage will be aligned, which were: input
sentence Si, the closest example of the source Se and its
translation St.
Fig. 1. Hybrid Machine Translation System.
Aligning the input sentence Si and the closest example Se
is done while computing the edit distance in equation (1). This
is shown in example (4) (4a1) with (4a2) are aligned, in (5)
(5a1) with (5a2) are aligned and in (6) (6a1) with (6a2) are
aligned. Then the closest example Se with its translation St
Input Sentence
Pre-processing
Example Based
Compute ED
Highest Score
Example Based
Translated Sentence
Un-matched portions
of Input and
Example
Translation
Memory
Align
Un-matched portions of
Example and translated
sentence
Replace &
Add or
Substitute
New Sentence with
Un-translated segments
Translation Memory
Translate un-
translated
segements
Translated
Sentence

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 10, No. 1, 2019
200 | P a g e
www.ijacsa.thesai.org
will be aligned by using the Translation memory that was built
and as shown (4a2) with (4a3) are aligned, (5a2) aligned with
(5a3) and (6a2) aligned with (6a3). In the next stage the
unmatched fragments will be replaced and the matched
fragments will keep unchanged.
(4) a

- impaired function of the [1:liver ]

- impaired function of the [1: kidneys]

- [1:

]

(5) a- 1- [1:arthrosis]
2- [1:arthritis]
3-

[1:

]
(6) b -1-[1: nasal] congestion
2- [1:lung ] congestion
3- [1:

]

Recombination stage: After extracting the unsuitable
fragments in the previous stage the next purpose is to adjust
the resulted translation. This is done by adding or substituting
the fragments from the input sentence (Si) with the translation
equivalent sentence (St) [16]. From example (4) {

} need
to be replaced from (4a3) with {liver} from (4a1), from
example (5) {

} need to be replaced from (5a3) with {
arthrosis } from (5a1) and from example (6) {

} need to be
replaced from (

a3)

with { nasal} from

(6a1). And the results
will be the sentence in (7), (8) and (9).
(7) liver

(8) arthrosis
(9) nasal

During the aligning the alignment might not only one to
one align. If the input sentence (Si) has extra segments that
have no align to translation equivalent sentence (St) this
segments are added to the final resulted sentence but if there is
extra segments in the translation equivalent sentence (St) they
will be deleted from the final resulted sentence. After this step
the task is to translate the un-translated segments using two
methods. The first method is to use the translation memory to
get the translation of the un-translated segments. The second
method is to use Google translate as a statistical machine
translation to get the translation of the un-translated segments.
The final result of the translation using Translation memory
showed in (10), (11) and (12).
(10)

(11)

(12)

V. RESULTS AND DISCUSSION
As said before two datasets were used in the experiments
for each dataset four experiments were made to measure the
accuracy of the proposed system using bilingual evaluation
understudy (BLEU) matrix. The datasets were divided to one
word sentences, two word sentences and multiple word
sentences and for each the experiments were made. In the first
experiment Google translate was used as it is a statistical
machine translation [32] that is

widely known with its
robustness, good performance, and the fact that it does not
require manually crafted rules [33] to translate the input
sentences. In the second experiment EBMT was used from it
matching stage only was used and the closet translation was
gotten and takes it as the translation for the input sentence. In
the third experiment the translation memory was used in
recombination stage to translate the unmatched portions. In
the fourth experiment Google Translate was used in
recombination stage to translate the unmatched portions.
BLEU score was used to automatically evaluate the proposed
system. BLEU score captures the fluency of the translation.
The following tables (Table 2 and Table 3) where the four
experiments were made for the whole dataset shows the
accuracy over the two datasets and as shown when using the
proposed system that uses both Example Based Machine
Translation and Translation Memory the results where the best
over the other techniques.
Results in Table 4 and Fig. 2 also in Table 5 and Fig. 3
show that over one word translation, two words translation
and multi-words translation the proposed approach achieved
the highest score over the four experiments and using Google
translate to translate the un-matched portions shows a very
bad score. Also as shown when the input sentence is from
multi-words the score increased.
As shown in Table 6 most of machine translation systems
in medical domain used Statistical machine translation
technique that will cause little accuracy with the dataset used
because of the Arabic morphology and the size of the corpus.
Their datasets were from systematic, clinical descriptions,
queries where they are from hospitals data but the core of the
used dataset were from internal medicine publications that are
used daily by patients and may contain complex data that need
translation.
TABLE II. SYSTEMS ACCURACIES FOR THE FIRST DATASET
System
BLEU
Google Translate
53.56
EBMT
48.86
EBMT+ Translation Memory
77.17
EBMT+ Google
73.07
TABLE III. SYSTEMS ACCURACIES FOR THE SECOND DATASET
System
BLEU
Google Translate
51.06
EBMT
50.82
EBMT+ Translation Memory
63.85
EBMT+ Google
61.43

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 10, No. 1, 2019
201 | P a g e
www.ijacsa.thesai.org
TABLE IV. SYSTEMS ACCURACIES FOR THE FIRST DATASET FOR
DIFFERENT INOUTE SIZE
System
Accuracy
Google
Translate
EBMT
EBMT+
Translation
Memory
EBMT
+
Google
1 word
translation
51.42
41.52
66.02
48.41
2 words
translation
51.32
47.36
59.21
19.89
Multi words
translation
54.23
52.99
80.93
74.47
Fig. 2. Comparison between Systems Accuracies for the First Dataset for
Different Input Size in the First Dataset.
TABLE V. COMPARISON WITH OTHER SYSTEMS
Reference number
Technique
Dataset type
The proposed system
18
statistical machine translation system
609 systematic reviews from English
to French
EBMT+ Translation memory
(translation system)
And the dataset used is using internal
medicine publications and Worldwide
Arabic Medical Translation Guide
Common Medical Terms sorted by
Arabic which is an English-Arabic
medical dictionary
Translating from English to Arabic
19
statistical machine translation system
English to Hindi
20
statistical machine translation system
clinical term descriptions from
Spanish to Brazilian Portuguese
21
Dictionary based machine translation
technique and Statistical machine
translation technique
query terms in medical domain from
English to German and vice versa
22
statistical machine translation system
medical data from English to Polish
and vice versa
23
statistical machine translation system
medical reports from English to
Tamil
2
Neural networks
medical data from English to Polish
and vice versa
24
Neural networks
French and German to English
17
Example based machine translation
technique matching stage
the internal medicine publications for
internal diseases
12
Transfer Approach
50 titles from the computer science
domain for training
66 real thesis titles from the computer
science domain for testing
Dataset is Medical Text.
Using Eaxmple Based technique with
Translation Memory
26
Rule-based transfer machine
translation technique
100 phrases and sentences from both
English and Arabic versions of
agricultural expert systems at CLAES
27
Rule-based
translate Greek to Greek Sign
language.
28
rule based approach.
well-structured English sentences into
well-structured Arabic sentences
3
Grammar parser and example based
machine translation technique
well-structured English sentences into
well-structured Arabic sentences
0
10
20
30
40
50
60
70
80
90
Google Translate
EBMT
EBMT+Transltion
Memory
EBMT+Google
BLEU Accuracy

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 10, No. 1, 2019
202 | P a g e
www.ijacsa.thesai.org
TABLE VII. SYSTEMS ACCURACIES FOR THE SECOND DATASET FOR
DIFFERENT INOUTE SIZE
System
Accuracy
Google
Translate
EBMT
EBMT+
Translation
Memory
EBMT
+
Google
1 word
translation
35.33
50.82

.20
14.31
2 words
translation
44.71
50.70
28.46
16.79
Multi words
translation
54.09
52.99
72.49
53.56
Fig. 3. Comparison between Systems Accuracies for the Second Dataset for
Different Input Size in the First Dataset.
Also as shown that when translation from English to
Arabic but not in medical domain mot of them used Rule
based technique where they analyze the English data in terms
of morphology, syntactic and semantic which is not necessary
in medical domain.
VI. CONCLUSION AND FUTURE WORK
A hybrid machine translation system using Example based
machine translation technique and Translation memory was
introduced in this paper to translate English medical terms to
Arabic medical terms in comparison with using Google
translate only to translate, Example based machine translation
system using matching stage only and finally with a hybrid
system using Example based machine translation technique
and Google Translate.
The system that used Example based machine translation
technique with a Translation memory achieved the highest
score in comparison with the other three experiments and this
because Translation memory that was used stores the
translation of each medical term then when using it to translate
the unmatched portions of the input sentence (Si) that were
added to the translated text (St) of the closest sentence (Se)
from the database in the recombination stage translation of
the unmatched portions to the right Arabic medical term will
be ensured. For the first dataset the proposed system achieved
77.17 % and for the second dataset 63.85%. Google translate
translates some of medical terms according to its English
meaning not according to its medical meaning. Also the result
from matching stage produces sentences with unmatched
words between the input sentence and the closest sentence
from the database. Using Google translate also with Example
based machine translation translates the some of the
unmatched portions according to its English meaning not its
medical meaning.
However, using one word translation, two words
translation and multi-words translation datasets achieved high
score for our system but the multi-words translation dataset
achieved the highest accuracy which is 80.93 % for the first
dataset and 72.49% for the second dataset. The reason for that
is because the training dataset contains multi-words sentences
more the one word sentences and also more than two words
sentences.
Adjusting the final result according to the morphology of
the Arabic language could make the resulted translation more
accurate.
REFERENCES
[1]           
       
International Journal of Computer Applications 121, no. 23 ,2015.
[2]      -based machine translation for
medical text domain. based on european medicines agency leaflet
 Procedia Computer Science, 64, pp.2-9, 2015.
[3] Alawneh, Mouiad, Nazlia Omar, T. Sembok, H. Almuhtaseb, and C.
Mellish., “Machine  International
Conference on Biomedical Engineering and Technology, 2011.
[4]      -specific expressions
.
[5] Neubig, Graham, et al., -reliability speech translation in
   The First Workshop on Natural Language
Processing for Medical and Healthcare Fields. 2013.
[6]            
      
    , In: Ninth Workshop on Statistical
Machine Translation, Baltimore, MD, USA Association for
Computational Linguistics, p.221-228, 2014 .
[7] Alsohybe, Nabeel T., Neama Abdulaziz Dahan, and Fadl Mutaher Ba-
  -translation history and evolution: survey for Arabic-
 arXiv preprint arXiv:1709.04685 ,2017.
[8]        
devices improving system of translating languages: what about their
        
Brazilian journal of cardiovascular surgery, 30(6), pp.664-667, 2015.
[9] Yepes, Antonio Jimeno, Elise Prieur-Gaston, and Aurélie Névéol,
 data to create parallel corpora for
 1, no.
1: 146, , 2013.
[10]  
International Journal of Science, Engineering and Technology
Research, 2(3), pp.pp-716, 2013.
[11] 
[12] Shaalan, Khaled, Ahmed Rafea, Azza Abdel Moneim, and Hoda
Baraka, "Machine translation of English noun phrases into
Arabic.", International Journal of Computer Processing of
Oriental Languages 17, no. 02, pp: 121-134, 2004.
[13] Gupta, Somya. "A survey of data driven machine Translation." Diss,
Indian Institute of ,2010.
[14] Costa-  
and comparison of rule-based and statistical Catalan-Spanish machine
translation systems Computing and informatics, 31(2), pp.245-270,
2012.
0
10
20
30
40
50
60
70
80
Google Translate
EBMT
EBMT+Translation
Memory
EBMT+Google
BLEU Accuracy

(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 10, No. 1, 2019
203 | P a g e
www.ijacsa.thesai.org
[15] Dandapat, S., Morrissey, S., Kumar Naskar, S. and Somers, H.,
  -based machine translation using
010.
[16] Seljan, Sanja, and Damir Pavuna. "Translation memory database in the
translation process." In Proceedings of the 17th International Conference
on Information and Intelligent Systems IIS 2006, pp. 327-332. Croatia,

[17] Rana Ehab   Example-based machine
translation: matching stage using internal medicine p 
International Conference on Software and Information Engineering
ICSIE, pp. 131-135, 2018.
[18] Névéol, A., Zweigenbaum, P., Max, A., Yvon, F., Ivanishcheva, Y. and
     ystematic reviews into
French, t 15(526), ,p.366K , 2013.
[19]     
      , International Journal of
Pure and Applied Mathematics, vol 118, no. 20, pp. 1649-1655, 2018.
[20] Renato, A., Castaño, J., Williams, M.D.P.A., Berinsky, H., Gambarte,
M.L., Park, H.J., Pérez-   A machine
translation approach for medical t HEALTHINF , pp. 369-378,
2018.
[21]  system description for
medical t    
on Statistical Machine Translation, pp. 229-232, 2014.
[22]      “Polish-English statistical machine
           
Internet Systems, Springer,Cham, pp. 169-179, 2015.
[23]            -
sensitive Machine Translation of Medical reports from English to
        
119, no. 16, pp. 297-304, 2018.
[24] Mikel Artetxe, Gorka Labaka, Eneko Agirre, and Kyunghyun Cho,
Unsupervised neural machine translation. In International
Conference on Learning Representations (ICLR), 2018.
[25] Amer, E. and Abd- wikipedia be a reliable source for
translation? testing wikipedia cross lingual coverage of medical
d ng (IOSR-JCE), Volume
18, Issue 3, PP 16-22, 2016.
[26] Shaalan, Khaled, Ashraf Hendam, and Ahmed Rafea, "An English-
Arabic bi-directional machine translation tool in the agriculture
domain.", In International Conference on Intelligent Information
Processing, pp. 281-290. Springer, Berlin, Heidelberg, 2010.
[27] Kouremenos, Dimitrios, Klimis Ntalianis, and Stefanos Kollias, "A
novel rule based machine translation scheme from Greek to Greek Sign
Language: Production of different types of large corpora and Language
Models evaluation", Computer Speech & Language 51, pp.110-135,
2018.
[28] Al-Taani, Ahmad T., and Zeyad M. Hailat, "A direct English-Arabic
machine translation system.", Information Technology Journal 4, no. 3
,pp: 256-261,2005.
[29]           
  , Procedia Engineering, 29, pp.3017-3022,
2012.
[30]        machine translation: a
survey, artificial intelligence r 42(4), pp.549-572, 2014.
[31] Artetxe, Mikel, Gorka Labaka, and Kepa Sarasola, "Building hybrid
machine translation systems by using an EBMT preprocessor to create
partial translations", In Proceedings of the 18th Annual Conference of
the European Association for Machine Translation, 2015.
[32] Costa-     Machine translation in medicine. a
quality analysis of statistical machine translation in the medical
  Conference on Advanced Research in Scientific Areas
(ARSA-2012), 2012.
[33]          
machine translation for biomedical tex      AMIA
Annual Symposium Proceedings, vol. 2011, p. 1290. American Medical
Informatics Association, 2011.