Model Transformation by Example with Statistical Machine Translation
Karima Berramla
1,3 a
, El Abbassia Deba
1 b
, Jiechen Wu
2 c
, Houari Sahraoui
2 d
and Abou El Hassen Benyamina
1 e
1
LAPECI Laboratory, Oran 1 University Ahmed Ben Bella, Algeria
2
DIRO, Universit
´
e de Montr
´
eal, Canada
3
Ain Temouchent University Center, Algeria
berramla.karima@edu.univ-oran1.dz, abbassia.deba@univ-oran1.dz, jiechen.wu@umontreal.ca,
sahraouh@iro.umontreal.ca, benyanabou@yahoo.fr
Keywords:
Model Transformation, Model Transformation by Example, Learning System, SMT System, IBM1 Model.
Abstract:
In the last decade, Model-Driven Engineering (MDE) has experienced rapid growth in the software devel-
opment community. In this context, model transformation occupies an important place that automates the
transitions between development steps during the application production. To implement this transformation
process, we require mastering languages and tools, but more importantly the semantic equivalence between
the involved input and output metamodels. This knowledge is in general difficult to acquire, which makes
transformation writing complex, time-consuming, and error-prone. In this paper, we propose a new model
transformation by example approach to simplify model transformations, using Statistical Machine Transla-
tion (SMT). Our approach exploits the power of SMT by converting models in natural language texts and by
processing them using models trained with IBM1 model.
1 INTRODUCTION
Model transformation is not a new activity. It has
been recognized as an essential component since the
appearance of application development techniques,
but it has essentially been implemented using gen-
eral purpose programming languages. The emergence
of the Model Driven architecture (MDA) (Kleppe
et al., 2003) paradigm has given another perspective
to model transformation by the introduction of dedi-
cated languages and tools.
Although dedicated languages and tools advanced
considerably the state of the practice, defining model
transformations still requires a lot of effort and time.
The current challenge in MDE is not only (i) how to
choose the best technique through which we define
and generate models and metamodels in a simple way
but also (ii) how to automate transformation process,
to alleviate the burden on developers who has to deal
with a variety of domain specific languages in which
a
https://orcid.org/0000-0002-2847-4895
b
https://orcid.org/0000-0003-2948-2093
c
https://orcid.org/0000-0002-4011-0859
d
https://orcid.org/0000-0001-6304-9926
e
https://orcid.org/0000-0003-4778-0123
models are expressed.
In this paper, we propose a new approach based
on a statistical machine translation (SMT) system for
model transformation by example. This approach
has several advantages. Firstly, it allows transform-
ing models without using a transformation language
that requires mastering its syntax and the semantic
between source and target metamodels. Secondly, it
reuses SMT, a technique with a proven track record
in natural language translation. Although SMT is
mainly used for natural language translation, this
technique was already used to solve some domain-
specific transformation problems, but in an ad-hoc
manner. An interesting example of such a usage is
the work by Alghawanmeh et al. (Al-Ghawanmeh
and Sma
¨
ıli, 2017), which translates Arab vocal im-
provisation to instrumental melodic accompaniment.
We evaluated our approach on varous transformation
examples. The results show that SMT-based model
transformation can be used successfully with a rea-
sonable number of examples.
The remainder of this paper is organized as fol-
lows. Section 2 defines the crucial problem of model
transformations. The proposed approach based on a
statistical machine translation system is presented in
Section 3 and a case study illustrates it in the sec-
Figure 1: Model transformation problems.
tion 4. Section 5 describes the evaluation of the ap-
proach. In Section 6 the related work is analyzed and
discussed. Section 7 concludes the work by focusing
on the advantages and limitations of the proposed ap-
proach.
2 PROBLEM STATEMENT
In MDE context, one founding assumption is that au-
tomated model transformation reduces the cost and
time of development and facilitates the development
through code generation. Model transformations also
allow providing intermediate bridges between the dif-
ferent development phases of an application. Usually,
transformation programs express the automatic trans-
lation of a model to another model according to a set
of transformation rules. Writing these rules requires
on the one hand, to have a good knowledge about the
semantics of each input and output metamodels and
on the other hand, to have a good knowledge of a
specific language in order to implement these trans-
formation rules. In this paper, we are interested in
the challenge of defining transformations with a min-
imum of domain knowledge. To address this chal-
lenge, we formulate the following research questions,
illustrated through the figure 1:
RQ1: Can we define the relationships between input
and output models using natural language?
RQ2: Can we generate the output models automati-
cally without using transformation rules that require
to use of a specific language?
RQ3: Can we use the natural language to define our
models or/and metamodels without using a specific
language in MDE context?
RQ4: Is there an optimal technique to model systems
without mastering a specific language or a modeling
tool?
When the model transformation is based on the spec-
ification of the correspondence rules between the el-
ements of the source and the target metamodels, the
domain experts have to implement it manually. To cir-
cumvent this burden, different research contributions
investigate the idea of deriving specification or con-
crete transformation from example through ad-hoc or
machine learning algorithms. The proposed solutions
are generally known as Model Transformation By Ex-
ample (MTBE). MTBE approaches use with differ-
ent techniques such as genetic programming (Baki
and Sahraoui, 2016) and Ad-hoc Algorithm (Varr
´
o,
2006).All these approaches attempt to generalize the
correspondence between source and target model el-
ements in the form of transformation rules written in
a specification or implementation language. Learning
complex structure such as rules is, however, a com-
plex problem.
Another perspective on transformation automa-
tion is to reuse the large body of knowledge on natu-
ral language translation, which is a very similar prob-
lem. In particular, statistical machine translation can
be adapted to MTBE. This technique is based on the
statistical models whose parameters are derived us-
ing a parallel (or bilingual) corpus in the learning
phase. Generally, another corpus is also used, that
is known as monolingual corpus, to define the word
resemblance in target language (Koehn, 2009).
We believe that SMT is suitable to our problem
because the translation is implemented through the
statistical training using a set of examples without re-
quiring transformation languages.
3 MTBE USING SMT
Our objective is to define a transformation mecha-
nism that is based on a corpus of transformation ex-
amples(TE). Our approach can be divided into three-
steps: (1) data preparation to map transformation ex-
amples into training data, (2) SMT training and (3) the
use of the trained SMT to actually transform models.
In the next, we explain these steps in detail.
3.1 Data Preparation
This step is a crucial part of our approach since
it allows to prepare the data that will be used to
train the statistical machine translation models. Data
preparation is done into two sub-steps: (1) trans-
forming source and target models into textual repre-
sentations, i.e., natural language sentences, and (2)
aligning source and target sentences that describe
semantically-equivalent source and target elements.
Figure 2 summarizes the data preparation step.
Figure 2: Corpus preparation process.
As in MDE, a model is an instance of its metamodel,
it can be decomposed into elements according to the
metamodel structure and constraints. Let us consider
the transformation problem of UML class diagrams
to relational schemas. For example, in a given UML
class diagram, one element can be a class ”Student”
with its attribute ”name”. For each element, we gen-
erate one or more sentences in a natural language.
Source models are translated into French, e.g., ”la
classe Student poss
`
ede attribut name”, which means
”class student has attribute name”. Target models
are translated into English. For example, in a rela-
tional schema corresponding to the considered class
diagram, we can find a table ”Student” with a col-
umn ”name”. This model element will be translated
into English as ”Table Student has column name”.
The translation of model elements into natural lan-
guage sentences can be done manually, as it can
be automated as a simple model-to-text transforma-
tion.
Observation 1: Ambiguity Problem
To solve the problem of ambiguity, we have
proposed a parallel corpus for each transforma-
tion example by considering the source meta-
model as the source language and the target
metamodel as also the target language.
Once the source and the target models are translated
into natural language sentences, the next step con-
sists in aligning source and target sentences that de-
scribe semantically-equivalent elements. Here again,
the alignment can be automatic if the input and the
output models are created simultaneously and auto-
matically (in our case, this process is done by using
Acceleo language for following example). Otherwise,
it can be done manually.
3.2 SMT Training
After the data preparation, the set of all aligned sen-
tences of all the pair examples define the parallel cor-
pus, which will be used later to train the translation
model. Additionally, the sentences of all the target
model examples define the monolingual corpus that
will be used to derive the language model.
SMT training is based on the parallel and mono-
lingual corpora of the translation system to be used. In
the following, we explain the basic concepts of SMT
system and we present the translation and the lan-
guage models that are used in this step. Figure 4 de-
scribes the training process of our SMT system. The
first model to train is the translation model. Accord-
ing to (Brown et al., 1993), a sentence f in a source
language has a possible translation to e in a target lan-
guage according to the probability P(e| f ). Starting
from the theorem of Bayes on the pair of sentences
(s,t), the probability P(e| f ) is computed according to
the following equation 1.
argmaxP(e| f ) = argmax
P( f |e).P(e)
P( f )
. (1)
The main objective of statistical translation is to find
the best translation ˆe, i.e. the value that maximizes
P(e| f ). In addition, the probability P( f ) of the source
sentence f has no influence on the result of this equa-
tion, because it is fixed. Sequentially, SMT transla-
tion process is defined by this simplified equation 2
that maximizes the probability P(e| f ) according to re-
alized observations in a parallel corpus.
ˆe = argmaxP(e| f ) = argmaxP( f |e).P(e). (2)
Figure 3: Training process.
P( f |e) defines a translation model based on inter-
lingual probabilities and P(e) explains a language
model to evaluate the probability of a word-sequence.
This equation and the following ones are explained
through case study in Section 4. In the following, we
detail the translation and language models.
3.2.1 Translation Model
The translation model defines the probability that a
word or word-sequence in a source language will be
translated into one or more words in a target language.
This model compute the probability of a target sen-
tence e
I
1
= e
1
... e
I
to be associated to a sentence from
the source language f
J
1
= f
1
... f
J
. In our work, we ex-
perimented with IBM1 model. Its translation model
is defined according to the following equation (Brown
et al., 1993).
P( f
J
1
|e
I
1
) =
ε
(I + 1)
J
I
a
1
=0
...
I
a
J
=0
J
j=1
t( f
j
|e
a j
). (3)
Where
I
a
1
=0
..
I
a
J
=0
defines the possible alignment and
J
j=1
t( f
j
|e
a j
) explains the translation probability of f
j
to e
a j
apply to all words in the source sentence.
3.2.2 Language Model
The objective of SMT is not only to produce a word-
by-word translation sequence as output, but also to
guarantee the grammatical reasonableness of the re-
sults in the target language. To achieve this objective,
many researchers propose to adapt language model
for machine translation system. This model is defined
by the following equation (Brown et al., 1993).
P(e
I
1
)
I
i=1
P(e
i
|e
in+1
, ...e
i1
). (4)
3.3 Model Transformation with SMT
In this sub-section, we describe our model transfor-
mation process using SMT system. Figure 4 gives
an overview of this process. Firstly, a model to be
transformed is translated into French sentences as de-
scribed in the data preparation process. Then, we use
the trained translation and language models to pro-
duce an equivalent English set of sentences describ-
ing the target model. Finally, this textual description
is mapper to a fully fledged target model according to
the associated metamodel.
The SMT translation is performed in two steps: a
decoding phase and a sentence and a model genera-
tion phase.
Figure 4: Model transformation process with SMT system.
3.3.1 Decoder
The decoder is the main component of the SMT sys-
tem. It consists in providing an output text from a
source text based on translation and language models.
This treatment is defined by the following equation
(Brown et al., 1993) and illustrated in Section 4:
ˆe = argmaxP( f
J
1
|e
I
1
)P(e
I
1
) = P( f
J
1
, ˆa
J
1
| f
I
1
)P(e
I
1
). (5)
P( f
J
1
, ˆa
J
1
| f
I
1
) defines the probability of the best transla-
tion that is based on the highest alignment probability
(in equation 5 this alignment is presented by ˆa
J
1
| f
I
1
).
3.3.2 Sentence and Model Generation
After the decoding using word-by-word translation,
the sentence generator assemble the translated words
into sentences according to the language model. From
these sentences, we generate the target model accord-
ing to the data preparation process.
4 ILLUSTRATIVE CASE STUDY
To better illustrate the proposed approach, we dis-
cuss UML class diagram to relational schema. This
case has been used in several papers in order to evalu-
ate model transformation approaches (see for example
(Baki and Sahraoui, 2016)).
Figure 5: Source and target metamodels.
4.1 Data Preparation
In this step, we are interested to define the parallel
and monolingual corpora in an automatic way by exe-
cuting a model-to-text transformations written in Ac-
celeo language. The generated sentences are apired
according to the paired model elements. Figure 6
shows an example of source-target model pair to-
gether with an excerpt of the paired sentences set.
Class 1
Class 2 Class 3
Attribut C1
Id_Class 2
Id_Class 3
Class 4
Id_Class 4
Attribut C4
1..*1..1
Class 2 Class 3
Id_2
Id_3
Class 4
Id_4
Attribut C4
1..*1..1
F 01
la classe 2
hérite de la classe
the table 2
contains
Attribut Id_2 de classe 2 a une valeur unique
Column Id_2 of Table 2 is a primary
………………………………………….
Class 6
Id_Class 6
Attribut C6
Class-Associa 5
Attribut C5
1..*
Class 6
Id_6
Attribut C6
Class-Associa 5
Attribut C5
Source Model
F 02
Classe_AssociationNM 5liée à
la classe
Table 5 has Foreign
-
primary
……………………………………..
1..*
Class 4
Id_Class 4
Attribut C4
Table 4Table 4
Id_Class 2Id_Class 2
Table 3Table 3
Id_Class 2Id_Class 2
Attribut C1
Table 2Table 2
Id_Class 2Id_Class 2
Attribut C1
Class 4
Attribut C4
Table 4Table 4
Id_4Id_4
Table 3Table 3
Id_3Id_3
Attribut C1
Table 2Table 2
Id_2Id_2
Attribut C1
hérite de la classe
1
contains
-all-columns of the table 1
Attribut Id_2 de classe 2 a une valeur unique
Column Id_2 of Table 2 is a primary
key
………………………………………….
Id_3Id_3
Id_Class 2Id_Class 2
Attribut C1
Table 6Table 6
Id_Class 2Id_Class 2
Attribut C1
Table 5Table 5
Id_6Id_6
Attribut C5
Id_4Id_4
Attribut C4
Table 6Table 6
Id_6Id_6
Attribut C6
Target Model
Classe_AssociationNM 5liée à
la classe
6.
Table 5 has Foreign
-Key-which-is
primary
-key-of the table 6
……………………………………..
Id_2Id_2
Id_3Id_3
Figure 6: Data preparation process for UMLCD2RS.
4.2 SMT Training
In this step, we describe the computation process of
both translation and language models.
Translation Model. In general, the translation
model gives the probability of the word translation
from a source language to another word in a target
language. Based on this probability, we generate the
target words that will be used then as a data to cre-
ate the target model. In the training phase, we cal-
culate the parameters of Equation 1 that will be use-
ful in the test phase in order to compute the probabil-
ity of each word translation. The calculation of these
IBM1 model parameters is based on the Expectation-
Maximization (EM) algorithm. This algorithm allows
to maximize the reasonableness of the parameters to
be learned (Dempster et al., 1977). The following ta-
ble describes some values of translation table that de-
fines IBM1 parameters.
Table 1: Translation table values.
P(f|e) value P(f|e) value
P(attribut|column) 0.999 P(est|column) 0.007
P(association1n
|foreign-key)
0.499 P(poss
`
ede
|foreign-key)
1e-12
P(classe |has) 0.015 P(poss
`
ede |has) 0.999
P(valeur
|primary)
0.497 P(unique |key) 0.497
P(classe|table) 0.836 P(unique|table) 1e-12
P(la|has) 0.103 P(la|the) 0.999
Language Model. In SMT systems, the language
model is used to compute the probability of P(e)
for a word-sequence t in a target language that e =
m
1
, m
2
, m
3
....m
n
. In our case, we used a bi-gram lan-
guage model that follows this probability (Brunning,
2010):
P(m
i
|m
i1
) =
c(m
i1
m
i
)
c(m
i1
)
(6)
Where c(m
i1
m
i
) defines the number of this word se-
quence m
i1
m
i
and c(m
i1
) explains the number of
m
i1
word. The following table describes the lan-
guage model values of UMLCD2RS example.
Table 2: Language model values.
P(m
i
| m
i1
) value P(m
i
| m
i1
) value
P(table|the) 0.992 P(there|the) 0.00
P(y|table) 0.349 P(column|table) 0.00
P(primary|a) 0.833 P(key|a) 0.00
P(has|x) 0.342 P(column|x) 0.00
P(has|primary) 0.000 P(key|primary) 0.97
P(table|the) 0.992 P(another-defined|the) 0.00
4.3 Transformation with SMT
To illustrate our translation, we select a set of sen-
tences written in natural language that define class3,
class4 and relationship between them in the source
model of figure 6.
Table 3: Comparative study of model transformation techniques.
Authors Input Elements Meta-modeling Tools Used Method Output Elements
Our Paper Sentences/Models Natural language SMT system Sentences/Models
(Hammoudi et al., 2014) ST-Metamodels Ecore Language Matching tools mappings
(Berramla et al., 2017b) ST-Metamodels Ecore Language Matching Alg Tr rules(A/C)
2
(Berramla et al., 2017a) ST-Models Ecore Language FCA-Matching Tr rules (A/C)
2
(Kessentini et al., 2012) Bloks /// PSO and SA Models
Elt1: attribut Id
3
de classe 3 a une valeur unique.
Elt2: attribut Id
4
de classe 4 a une valeur unique.
Elt3: la classe 4 poss
`
ede attribut C4.
Elt4: la classe 4 poss
`
ede association1n avec la classe
3.
Decoder. In this step, the decoder uses the param-
eters of IBM1 model
1
to find the best translation of
source text from bilingual and monolingual corpora.
The calculated parameters are shown in table 1 and
2. Once previous tables (1 and 2) are created, the de-
coder can generate the target sentences from source
sentences based on the following equation.
ˆe = P( f
J
1
, ˆa
J
1
| f
I
1
)P(e
I
1
) =
J
j=1
ˆ
t( f
j
|e
a j
)
I
i=1
P(e
i
|e
i1
). (7)
IBM1 model is not based on the alignment where
P( f
J
1
, ˆa
J
1
| f
I
1
) =
ˆ
t( f
j
|e
a j
) that explains the best trans-
lation value which is selected from the previous table.
Sentence and Model Generation. Once the decod-
ing phase is done, the extraction of target sentences
can be executed based on this calculated probability
argmaxP( f
J
1
|e
I
1
)P(e
I
1
). The output sentences of the
input sentences (see Figure 6 and its definition in nat-
ural language in this sub-section) are defined as fol-
lows:
Translation of Elt1: column Id
3
of table 3 is a pri-
mary key.
Translation of Elt2: column Id
4
of table 4 is a pri-
mary key.
Translation of Elt3: the table 4 has column C4.
Translation of Elt4: the table 3 has foreign-key that-
is-defined-as-primary-key-in the table 3. To have use-
ful MDE artifacts, it is necessary to map the sentences
generated by SMT system into models conform to re-
lational schema metamodel. In our case, we create
the models manually from their correspondence sen-
tences.
1
https://www.nltk.org/ modules/nltk/translate/ibm1.
html
5 EXPERIMENTATION AND
DISCUSSION
The objective of this section is to evaluate the pro-
posed approach with six transformation examples
from various domains. For each transformation we
collected/defined a set of pairs of model examples.
The models were translated into natural language
texts. Table 4 shows the vocabulary volume (sen-
tences and words) for parallel corpus of each trans-
formation example (in our case, the monolingual cor-
pus size generally presents the half (1/2) of parallel
corpus for each example).
Table 4: Information about used transformation examples.
Transformation Examples Parallel Corpus
Example NO
Exples
NO
Sents
NO
Words
UMLCD2RS 10 154 1073
Book2Publication 10 66 688
Family2Person 10 88 1016
Ecore2Coq 10 392 3071
AADL2TA 05 118 1217
UMLSM2PetriNet 15 388 3899
We evaluate the precision and recall metrics as
adapted to model transformation in (Hammoudi et al.,
2014). Both metrics are calculated in terms of words
found by the automated SMT translation (W
auto
) w.r.t.
the words expected (W
expert
), i.e, defined by the expert
who produced the model example pair. We define the
set W
positive
=W
auto
W
expert
, as the set of words that
are obtained by SMT translation and are also included
in the expert’s words.
Figure 7: Precision and Recall of transformation examples.
In this context, we applied the trained translation
model on the examples used for training. Figure 7
summarizes the results obtained for the precision and
recall. Our SMT system based on IBM1 model gives
perfect precision 100% for all transformation exam-
ples. For the recall our SMT system exhibits also per-
fect values 100% for all transformation examples ex-
cept for AADL2TA, which are considered as a very
less score in our SMT system (these values are de-
fined between 88% and 93%).
These values reflect the difficulty of fully han-
dling transformation with complex semantic gaps be-
tween source and target models such as AADL2TA
example, but with well structured semantics such as
UMLCD2RS transformation we have obtained the
best values.
In our case, we generate the target models (in our
case, from target sentences are created the target mod-
els) without using a set of transformation rules but
only with a set of source and target model pairs. This
latter is used to specify the parallel and monolingual
corpora for our SMT system. In this case, we have
processed the following problems: RQ3 and RQ4 that
are presented in section 2 have been solved by speci-
fying the models in natural language using manual or
automatic preparation).
In model transformation context, all proposed
works are based on optimization technique or match-
ing tool in their main step, however,our approach
aims to ease the transformation by treating the chal-
lenge of RQ2 by using exactly SMT system to gen-
erate the output models without using a specific pro-
gram.
Observation 2: Using NL in MDE Context
The use of natural language facilitates meta-
modeling and model transformation processes.
These latter are done automatically without ma-
nipulating specific languages or tools.
Our approach makes it possible to solve the problem
RQ1 focusing on learning system based on parallel
and monolingual corpora using IBM1 model. In this
case, parallel and monolingual corpora are proposed
as a definition of transformation rules and the execu-
tion of IBM1 is defined as the executable of transfor-
mation rules.
6 RELATED WORK
The model transformation challenge is still how to
automate the model transformation process. Several
studies suggested to interpret this challenge into two
principal classes according to the type of input ele-
ment: the first class is based on metamodels and the
second one manipulates models to automate or semi-
automate transformation process.
First class is focused on the detection of relation-
ships between the input and the output metamodels
automatically. It is important to notify that some
works are focused on specifying semi-automatically
a transformation rules by using matching techniques.
For instance, this work (Hammoudi et al., 2014) de-
scribes the use of various matching tools and tech-
niques to automate model transformations but these
matching techniques are based also on metamodel
pretreatment and require understanding the basic in-
formation about the ontologies. Another example
that illustrates this idea is Berramla’s paper (Berramla
et al., 2017b) that is focused on matching algorithm
but it compares all metamodel elements without con-
sidering their types and uses metamodels as input data
without changing their structure. Both abstract and
concrete transformation rules are generated from sim-
ple mapping elements, however, the intervention of an
expert is obligatory for complex transformation ex-
amples to validate their generated mappings.
Second class is founded on the use of models
as input elements to automate model transformation.
This category contains also several proposed works
using different techniques. In what follows, we briefly
discuss on these prpoposed works. Some works are
based on MTBE using ad-hoc algorithms. One among
the first researches in model transformation by exam-
ple is the Varr
´
o ’s work (Varr
´
o, 2006) that uses as in-
put elements a set of source and target model pairs and
the prototype mapping between each model pair in or-
der to help the transformation designer the creation
of model transformation semi-automatically. Other
studies use artificial intelligence (AI) techniques. In
(Baki and Sahraoui, 2016) is based on genetic pro-
gramming to learn operational rules from source and
target model pairs and a set of transformation rules
by adding also the control of each rule during the their
generation. In (Kessentini et al., 2012) describe a new
approach to define how to integrate Particle Swarm
Optimization (PSO) and Simulated Annealing (SA)
in order to generate the output models automatically
without using a transformation program. The rest of
these studies integrate the mathematical methods to
define the model transformation by examples. For in-
stance, in (Berramla et al., 2017a) a new approach
is proposed defining the hybridization between the
model-matching and the formal concept analysis. The
first one is used to create the correspondences from
2
Transformation rules written in abstract level(or and in
concrete levels).
source and target models and the second one is ap-
plied to reduce the redundancy of transformation rule
creation.
From the table 3, we note that all proposed ap-
proaches are based on metamodeling tools or lan-
guages to have a best representation of their input el-
ements which require a little times and efforts to in-
terpret their input elements but with our approach, we
use only the natural language to describe the models
to be translated.
The main challenge of the most proposed works
in this context is how to create transformation rules
semi-automatically. This paper is focused on how
to translate easily a model to another model without
using tools or languages in both metamodeling and
transformation levels.
7 CONCLUSION AND FUTURE
WORK
During recent years, model transformation by ex-
ample has seen in several works such as (Baki and
Sahraoui, 2016; Varr
´
o, 2006) to define the model
transformation in a semi-automatic way but applying
the proposed techniques are limited to simple trans-
formation examples. In this paper, we proposed an
approach to define the model transformation automat-
ically using only a set of models and SMT system
more specifically IBM1 model in order to reduce the
costs and the time of software development.
The most important objective of this work is not
only to automate transformation process but also to
facilitate the modeling phase by using natural lan-
guage without basing on specific tools or/and lan-
guages that require good knowledge about them.
Once the modeling phase is executed, a set of sen-
tences written in two languages are builded from the
source and the target model pairs. From these sen-
tences, is established a parallel corpus which is useful
in the training phase to calculate a set of parameters
that has been also used in the test phase in order to
obtain a good translation with this system. Finally the
translated sentences permit to create automatically the
target models through the use of their metamodels.
This process changes from one transformation exam-
ple to an another according to the structure of target
metamodel.
Improvements can be planned as perspectives for
this work. First of all, to propose a generalization of
the transformation process that will make it possible
to translate any model into a set of sentences written
in natural language. Also, we can use other transla-
tion system such as phrase-based translation system.
ACKNOWLEDGEMENTS
This work has been funded in part by the europian
project PRIMA WaterMed 4.0, ”Efficient use and
management of conventional and non-conventional
water resources through smart technologies applied
to improve the quality and safety of Mediterranean
agriculture in semi-arid areas” and by MESRS, ”Min-
ist
`
ere de l’enseignement sup
`
erieur et de la recherche
scientifique”.
REFERENCES
Al-Ghawanmeh, F. and Sma
¨
ıli, K. (2017). Statistical ma-
chine translation from arab vocal improvisation to
instrumental melodic accompaniment. Journal of
the International Science and General Applications,
1(1):11–17.
Baki, I. and Sahraoui, H. (2016). Multi-step learning and
adaptive search for learning complex model transfor-
mations from examples. ACM Trans. Softw. Eng.
Methodol., 25(3):20:1–20:37.
Berramla, K., Deba, E. A., Benyamina, A., Touam, R.,
Brahimi, Y., and Benhamamouch, D. (2017a). Formal
concept analysis for specification of model transfor-
mations. In Embedded & Distributed Systems (EDiS),
pages 1–6. IEEE.
Berramla, K., Deba, E. A., Benyamina, A. E. H., and Ben-
hamamouch, D. (2017b). A Contribution to the Spec-
ification of Model Transformations with Metamodel
Matching Approach. IJISMD, 8(3):1–23.
Brown, P. F., Pietra, V. J. D., Pietra, S. A. D., and Mercer,
R. L. (1993). The mathematics of statistical machine
translation: Parameter estimation. Computational lin-
guistics, 19(2):263–311.
Brunning, J. J. J. (2010). Alignment models and algorithms
for statistical machine translation. PhD thesis, Uni-
versity of Cambridge.
Dempster, A. P., Laird, N. M., and Rubin, D. B. (1977).
Maximum likelihood from incomplete data via the em
algorithm. Journal of the royal statistical society. Se-
ries B (methodological), pages 1–38.
Hammoudi, S., Feki, J., and Lafi, L. (2014). Metamodel
matching techniques: Review, comparison and eval-
uation. International Journal of Information System
Modeling and Design, pages 1–27.
Kessentini, M., Sahraoui, H., Boukadoum, M., and Omar,
O. B. (2012). Search-based model transformation by
example. Software & Systems Modeling, 11(2):209–
226.
Kleppe, A. G., Warmer, J., Warmer, J. B., and Bast, W.
(2003). MDA explained: the model driven architec-
ture: practice and promise. Addison-Wesley Profes-
sional.
Koehn, P. (2009). Statistical machine translation. Cam-
bridge University Press.
Varr
´
o, D. (2006). Model transformation by example. In In-
ternational Conference on Model Driven Engineering
Languages and Systems, pages 410–424. Springer.