BPMN Model and Text Instructions Automatic Synchronization
Leonardo Guerreiro Azevedo
1,2
, Raphael de Almeida Rodrigues
1
and Kate Revoredo
1
1
Graduate Program in Informatics (PPGI), Federal University of the State of Rio de Janeiro (UNIRIO),
Av. Pasteur, 456, Urca, 22290-240, Rio de Janeiro, RJ, Brazil
2
IBM Research, IBM, Av. Pasteur 146, Botafogo, 22290-240, Rio de Janeiro, RJ, Brazil
Keywords:
BPM, BPMN, Natural Language Generation, Natural Language Processing.
Abstract:
The proper representation of business processes is important for its execution and understanding. BPMN is
the de facto standard notation for business process modeling. However, domain specialists, which are experts
in the business, do not have necessarily the modeling skills to easily read a BPMN model. Natural language is
easier to read for them. So, both model and text are necessary artifacts for a broad communication. However,
unilateral manual editions of them may result in inconsistencies. This research proposes a framework for
synchronizing BPMN model artifacts and its natural language text representation. It generates textual work
instructions from the model, and it updates the original model if the textual instructions are edited. The
framework was implemented using Java standard technology and evaluated through experiments. In the first
experiment, we showed the knowledge represented by the textual work instructions and the correspondent
process models are equivalent. Furthermore, in a second experiment, we showed our approach for maintaining
the texts and models consistent performed satisfactory, where we verified the equivalence of the two artifacts.
1 INTRODUCTION
Many companies maintain both process models and
textual work instructions to depict their processes
(van der Molen, 2011). The use of both represen-
tations is needed to address specific audience. Usu-
ally, domain experts are not qualified for reading pro-
cess models, due to lack of knowledge in the mod-
eling notation, and they rely on the textual descrip-
tions. On the other hand, modeling experts prefer
using the model representation which are easier for
them to read and understand. Companies face redun-
dant effort for updating both process knowledge rep-
resentation artifacts. This task is error-prone, and may
bring inconsistencies.
The scientific problem addressed by this work is
how to maintain both knowledge representation arti-
facts (texts and models) synchronized automatically,
ensuring their consistency. We aim at enabling mod-
elers to generate texts from models and domain ex-
perts to create or update formal models through tex-
tual descriptions writing and editions. As a result,
the proposal leverages the information potential of al-
ready existing text documents. It preserves process
verbalization techniques. It provides substantial sav-
ings, reducing manual efforts. It solves the inconsis-
tency model-text problem. It enables a quicker real-
ization of BPM-projects and their benefits.
This work proposes a round-trip framework for
automatically maintaining the consistency of pro-
cesses representations by reflecting to a process
model modifications done to the text and vice-versa.
The approach generates natural language text from
a process model, and updates the process model
from edited text. To perform these synchronizations,
the framework creates correlations among process
model’s elements and the text. So, if one change one
of the artifacts, it can use the framework to automati-
cally update the other.
Using our approach domain specialists do not
need to be trained in a specific business process mod-
eling notation or in process modeling discipline. They
can update a process model by editing the natural lan-
guage text derived from the original process model.
On the other hand, system analysts are relieved from
the time-intensive task of writing and updating natural
language texts that represents the models.
The remainder of this work is structured as fol-
lows. Section 2 presents the proposed round-trip
framework and its implementation. Section 3 presents
the evaluation. Section 4 presents the related works.
Finally, Section 5 presents the conclusions, proposals
of future work and existing framework’s limitations.
484
Guerreiro Azevedo, L., de Almeida Rodrigues, R. and Revoredo, K.
BPMN Model and Text Instructions Automatic Synchronization.
DOI: 10.5220/0006809604840491
In Proceedings of the 20th International Conference on Enterprise Information Systems (ICEIS 2018), pages 484-491
ISBN: 978-989-758-298-1
Copyright
c
2019 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
2 THE ROUND-TRIP
FRAMEWORK
This section presents the round-trip framework. Sec-
tion 2.1 presents an overview of the framework and
the scenarios it supports. Section 2.2 presents the trip
to generate text from business process models, while
Section 2.3 presents the trip to generate business pro-
cess models from text instructions.
2.1 Framework Overview and
Supported Scenarios
Figure 1 gives an abstract overview of the round-trip
framework. One can starts from models or texts, top
and bottom of the figure, respectively. The links be-
tween model’s elements and text’s sentences (mid-
dle of the figure) are created through pattern match-
ing and stored for synchronization queries executed
by Process Model to Natural Language (labeled as P)
and Natural Language to Process Model (labeled as
N) components. The pattern matching rules are de-
fined as templates.
Figure 1: Abstract overview of the round-trip framework.
The framework fits the following scenarios:
Generate Text from Process Model: The frame-
work receives a model as input, triggers the
Process Model to Natural Language component,
which queries the linked elements, and generates
the whole text as natural language sentences.
Generate Model from Process Textual Descrip-
tions: The framework receives a text as input,
triggers the Natural Language to Process Model
component, which queries the linked elements,
and generates the whole model.
Update A Model from Edited Text: The frame-
work’s Synchronization Component (labeled as S)
detects changes in the text, and reflects them to the
process model. This is done by triggering the Nat-
ural Language to Process Model component and,
thus, generating an updated version of the model.
This scenario only reflects changes, updating only
the process model’s elements that were altered.
Update a Text from Edited Model: The text up-
date is handle as a new text generation scenario,
i.e., the whole model is submitted as input to the
Process Model to Natural Language component
and the whole text is regenerated. The decision
to generate the whole text again instead of updat-
ing only sentences which had changes was based
on cost-benefit analysis, considering the perfor-
mance gain and development effort.
Round-Trip: This scenario is represented by a
generation cycle. For illustration purpose, con-
sider the process model as starting point. The
model is submitted as the framework’s input, and
the text is generated. Then, the text is manually
changed and the original process model is auto-
matically updated. In the same cycle, the up-
dated model is submitted again to assert whether
it matches the text description. At this point, the
user can go in both directions, generate text from
model or model from text.
The NLG and NLP core components are com-
posed by several ready-to-use building blocks (Frozen
spots) and define interfaces which must be imple-
mented to support specific languages (Hot spots).
Interfaces provide the flexibility to satisfy specific
needs. For instance, regarding NLG, each hot spot
can be implemented for a specific language (e.g., Ger-
man and Spanish) without the need to change any
component. Regarding NLP, the hot spots can be
implemented to override the default behaviour or to
offer support to different text formats (i.e., text pat-
terns). The architecture’s frozen spots are represented
by classes, while the hot spots are represented by in-
terfaces (Pree, 1994).
Figure 2 presents a package diagram of the
hot spots implementations developed for Por-
tuguese and English. They are suffixed as Re-
alizer since they realize the implementations of
GeneralLanguageCommon package interfaces. Each
language has its own specific implementation
(e.g., PortugueseLabelHelper class implements
ILabelHelper). So, PortugueseRealizer and
EnglishRealizer classes implement hot spots and
use frozen spots to accomplish necessary tasks.
BPMN Model and Text Instructions Automatic Synchronization
485
Figure 2: Implementation of the hot spots defined by the
architecture.
2.2 Model to Text: Natural Language
Generation from BPMN Process
Model
This section describes the component Process Model
to Natural Language, i.e., the NLG component.
The main challenge for generating text from pro-
cess models is to adequately analyze the existing
natural language fragments from the process model
elements, and to organize the information in a se-
quential fashion. The Model to Text component
of the framework was implemented following the
Three-Step NLG pipeline proposed by Reiter and
Dale (Ehud Reiter, 1997), which are:
Text Planning: Determines the information to be
communicated in the text, and specifies the order
this information should be conveyed.
Sentence Planning: Chooses specific words to ex-
press the information determined in the previous
step. It may aggregate messages and include pro-
nouns in the text to obtain smoothly text.
Sentence Realization: Transforms the messages
into grammatically correct sentences.
The pipeline was instantiated as presented in Fig-
ure 3 and detailed as follows:
Text Planing
Linguistic Information Extraction: Decom-
poses process model element’s labels using
a linguistic label analysis technique (Leopold
et al., 2013). For instance, it decomposes
the activity label Inform customer about prob-
lem into action (inform), business object (cus-
tomer), and addition (about problem).
Annotated RPST Generation: Creates Re-
fined Process Structure Tree (RPST) (Vanhat-
alo et al., 2009)) from the process model.
Text Structuring: Annotates the RPST tree’s
nodes with the linguistic information obtained
in the first step.
Sentence Planning
DSynT-Message Generation: Maps the anno-
tated RPST elements to a list of intermedi-
ate messages. Each sentence is stored as a
Deep-Syntactic Tree (DSynT) (Mel’
ˇ
cuk and
Polguere, 1987).
Message Refinement: Aggregates messages,
generates referring expressions (e.g., replace
the role analyst to he), and inserts discourse
marker (e.g., afterwards or subsequently) if
needed. It performs these refinements usually
on long sequences of tasks.
Realization
Surface Realization: Transforms the interme-
diate messages into grammatically correct sen-
tences. This is accomplished by systematically
mapping the generated DSynT to the corre-
sponding grammatical entities.
2.3 Text to Model: BPMN Process
Model Generation from Natural
Language Texts
This section describes the component Natural Lan-
guage to Process Model, i.e., the NLP component.
It uses NLP techniques to extract and parse texts
to identify BPMN process model elements and cre-
ate/update the model. It was implemented as pre-
sented in Figure 4 and detailed as follows:
Text Planning
Text Pattern Extraction: Infers the textual pat-
tern from the text received as input. It vali-
dates if the text received as input follows the
supported text patterns, thus saving both user’s
time and system processing resources.
Natural Language Processing: Analyzes the
text’s semantic and extracts relevant linguis-
tic information. It employs NLP techniques
to: remove stop words (e.g., articles and dis-
course markers, like Then, Afterwards); recog-
nize patterns (e.g., end of line); split the text
into sentences; split the sentences into words
array and classify words according to their re-
spective syntactic rule, i.e., it executes Part-Of-
Speech Tagging (POS), which consists of tag-
ging each word with its respective syntactic role
(e.g., plural noun, adverb).
Text Cleaning: Replaces referring expressions
and sentences aggregations (e.g., replaces he by
the role analyst).
Sentece Text Planning
ICEIS 2018 - 20th International Conference on Enterprise Information Systems
486
Figure 3: Extended pipeline approach for NLG.
Sentence Text Generation and Structuring:
Generates SentenceText objects for the original
RealizedText while preserving the description
structure, e.g., generates activities in the correct
sequence order and correlation among them.
Process Moderl Realization
Model Elements Generation: Generates the
model elements from the SentenceText objects.
The support of multiple text structures and for-
mats is this step challenge.
Process Model Structuring: Structures the gen-
erated elements according to the order which
they appear within the text and also respecting
sentences correlations. It takes advantage of
the link between the SentenceText and the as-
sociated model elements, and it benefits from
the initial sentences structuring. Each Sen-
tenceText is mapped only to one model ele-
ment, and it is just needed to traverse through
the sentences tree nodes, fetch its associated
model element and create the ProcessModel
object. With the ProcessModel object (which
is not notation specific), it is possible to trans-
late it to any specific notation and print out the
whole process model to the desired output (e.g.,
JSON, XML, BPMN, etc), which is done by the
next and final NLP’s pipeline component.
BPMN Model Generation: Generates a busi-
ness process model in a BPM notation. In this
work implementation, the used notation was
BPMN. It generates a JSON file that represents
the process model as a machine artifact. The
JSON file can then be used in a BPMN graphic
modeler tool to dispose all the elements visu-
ally. In particular, Signavio
1
online platform
was used to read the JSON file and output the
graphic process model to the user.
1
available at www.signavio.com
3 EVALUATION
This section presents the evaluation of the techniques
developed within the language-independent frame-
work.
The first experiment’s objective was Assess
whether the knowledge represented by the generated
process description can be considered equivalent to
the process model. It was guided by the following
research question (RQ1): “Is the knowledge repre-
sented by the natural language text, generated by the
framework, equivalent to the process model?”. Thus,
the focus were in determining how many answers
2
were within the equivalence range varying between
100 and 68%, how many were between 67 and 34%,
and, finally, how many were between 33 and 0%. The
expectation was the number of answers between 100
and 68% be higher than the sum of the other two
groups (participants who rate equivalence between 67
and 0%).
Figure 5 depicts the overall evaluation for this ex-
periment. As can be observed, 342 answers were
within the equivalence group ranging from 100% to
68%, which can be read as 74% of the participants
claim that the equivalence between both knowledge
representations vary from 100% to 68%
3
. We ar-
gue the chosen text structure can achieve the expected
results, which is to transmit the process knowledge
through a natural language representation. Based on
this, the answer for RQ1 is: The knowledge repre-
sented by the natural language text, generated by the
framework, can be considered equivalent to the pro-
cess model.
2
The questionnaire is available at
https://docs.google.com/forms/d/
1zvuuxojWUVWnyRpMsao-F1Yjygs-Dm2ebfrMjjjJqx4/
viewform (in Portuguese)
3
Each participant contributed with seven (7) answers.
Thus, if we divide the total number of answers by seven
(342/7) it gives us the average number of participants that
choose the same answer (48 participants).
BPMN Model and Text Instructions Automatic Synchronization
487
Figure 4: Text to model pipeline.
Figure 5: Subject’s answers distribution among equivalence
intervals.
The second experiment’s objective was Assess
whether the knowledge represented by the automati-
cally updated version of the process model, after been
synchronized by text-based changes, represents the
same knowledge as the manually updated textual de-
scription. It was guided by the following research
question (RQ2): “Is the knowledge represented by
the manually updated text equivalent to the auto-
matically updated process model?”. Thus, the fo-
cus was to determine how many answers were within
the range varying from Strongly Disagree to Strongly
Agree. The expectation was the number of answers
4
between Strongly Agree and Slightly Agree be higher
than the sum of the other groups (subjects who rate ac-
cordance with the synchronization strategy between
Neutral and Strongly Disagree). Thus, to enable a
better reading of the results, the answers for options
Strongly Agree and Slightly Agree were grouped into
one group. All the remaining answers were grouped
into a second group.
Figure 6 depicts the overall evaluation for this
question. As can be observed, 160 answers were
within the equivalence group ranging from Strongly
Agree and Slightly Agree, which can be read as 78%
of the subjects
5
claims the knowledge represented by
4
The questionnaire is available at
http://goo.gl/forms/JacmFbV3yTpPEfib2 (in Portuguese)
5
Each subject contributed with seven (14) answers.
the manually updated text is equivalent to the auto-
matically updated process model. Based on this, the
answer for RQ2 is: The knowledge represented by the
manually updated text can be considered equivalent
to the automatically updated process model.
Figure 6: Subject’s answers distribution among the avail-
able accordance options (grouped into two groups).
4 RELATED WORK
Several related works proposed similar approaches
to our framework. They can be grouped into three
(3) specific topics: (i) Text to model generation; (ii)
Model to textual description generation; and, (iii)
business process understandability.
4.1 Text to Model and Model to Text
Generation
We inspected related works where natural language
techniques was used to achieve BPM relevant goals
or areas where the creation of process models was the
focus. These approaches provided initial insights for
the construction of the NLG Core module architec-
ture). They are presented in Table 1 along with com-
parisons with our work.
Thus, if we divide the total number of answers by fourteen
(160/14), it gives us the average number of subjects that
choose the same answer (11 subjects)
ICEIS 2018 - 20th International Conference on Enterprise Information Systems
488
Table 1: Papers’ Comparison.
Paper and Objective Description Comparision
Leopold et al.
(Leopold et al.,
2012):
Generate natural
language texts from
BPMN models.
Approach that automatically transforms
BPMN process models into natural lan-
guage texts. It is based on Reiter
and Dale’s NLG pipeline (Ehud Reiter,
1997). Moreover, Leopold et al. pro-
posed a new approach that validates pro-
cess model through natural language gen-
eration (Leopold et al., 2014).
Our framework is capable of parsing pro-
cess models and generating textual pro-
cess descriptions from these models like
Leopold et al. approach. However, our ap-
proach is capable of generating the model
from the text and is not language specific.
Gonc¸alves et
al. (de AR Goncalves
et al., 2009):
Generate conceptual
models from natural
language texts.
A method for deriving conceptual mod-
els from text focusing on the derivation of
models from group stories. They provide a
prototype which handles Portuguese texts.
As opposed to our work, theirs offer no
support to any other language, it parses
only story like texts, considered a limited
set of BPMN elements and provides a sim-
ple evaluation. Furthermore, a couple of
their exhibits show that syntactical prob-
lems occur in some cases.
Ghose et al. (Ghose
et al., 2007):
Generate UML com-
pliant process mod-
els from natural lan-
guage texts.
Toolkit that uses a syntax parser to identify
verb-object phrases in the given text and
also scans the text for textual patterns.
The results are BPMN model snippets
rather than a fully connected model.
Kop and Mayr (Kop
et al., 2005):
Generate mod-
els from natural
language texts.
Proposed a procedure called KCPM (Kla-
genfurt Conceptual Predesign Model) and
developed a corresponding tool. It parses
textual input in German and fills instances
of a generic meta-model, the KCPM. Us-
ing the information stored in this meta-
model, an UML activity diagram and a
UML class diagram can be created.
As opposed to our work, the transforma-
tion from natural language input to the
aforementioned meta-model is not a fully
automated process and supports only Ger-
man written texts.
Sinha et al. (Sinha
and Paradkar, 2010):
Generate process
model from natural
language text.
They employ a linguistic analysis engine
based on the UIMA Framework, which en-
ables the constructions of linguistic analy-
sis systems by combining different blocks
into a pipeline.
Their work does not contain a full text and
model example, which does not allow a
comparison with ours. Besides, it is a text
type specific since it uses only use-case de-
scriptions.
Wang et al. (Wang
et al., 2009):
Generate BPMN
model from struc-
tured use cases.
Describe a procedure which creates a
BPMN diagram, given that data items,
tasks, resources (actors) and constraints are
identified in an input text document.
Compared to Wang et al. work, our ap-
proach does not require user-interaction
during the parsing execution. Ours sup-
ports 15 different BPMN elements, and
does not impose any restriction regard-
ing the number of gateways outgoing arcs.
They also do not present any process
model they use.
Friedrich et
al. (Friedrich et al.,
2011):
Generate model from
natural language
text.
Proposes an automatic procedure for gen-
erating BPMN models from natural lan-
guage text composed by an anaphora res-
olution component. The evaluation count
with more than 45 models from both,
academy and industry.
Their work generates a conceptual model
from natural language text and does not
generate natural language texts from the
models. It is capable of parsing only En-
glish texts and they do not present possi-
bilities to extend or adapt the technique to
fit specific needs.
4.2 Business Process Understandability
The field of process model understandability is dis-
cussed from different perspectives.
Mendling et al. show the number of arcs has an
important effect on the overall model understandabil-
ity (Mendling et al., 2007). Mendling et al. demon-
strates the impact of natural language in the activ-
ity labels for model comprehension (Mendling et al.,
2010). Zugal et al. investigated how far the cogni-
BPMN Model and Text Instructions Automatic Synchronization
489
tive inference process affects the model understand-
ing (Zugal et al., 2011). Leopold et al. build on these
insights trying to lower the overall burden of process
model comprehension (Leopold et al., 2014).
Prior work comparing process modeling nota-
tions can be roughly grouped into two categories:
(i) Graphical notation comparison; (ii)Textual versus
graphical notation comparison. The first is the most
prominent one. Several of these studies suggest ex-
pertise is the most relevant factor in comprehension
(Curtis et al., 1989), but there is no absolute better
or worse representation. In a previous work, we run
an experiment on process model understandability us-
ing textual work instructions and BPMN Models. The
results were instructions or process models do not in-
fluence process understandability for non-expert users
but do influence experienced users (Rodrigues et al.,
2015).
Haisjackl and Zugal compared declarative pro-
cess models against a text based notation using sub-
jects with experience in modeling declarative pro-
cess (Haisjackl and Zugal, 2014). Different from their
work, ours used imperative process models, involved
subjects with experience in process modeling varying
from none to expert, and presented a natural language
text simulating a human description of the process.
While the experiment described by Ottensooser et
al. compares model understanding with written use
cases (using the Cockburn format), our framework is
based on a more fluent and natural representation of
the text (Ottensooser et al., 2012). More specifically,
it was considered a natural language text which re-
quires no background knowledge of layout or specific
patterns. Also, our research focus on process under-
standing while Ottensooser et al. focus on domain
understanding through different representations. Nev-
ertheless, our work’s findings corroborates to Otten-
sooser’s results which indicate there is no significant
superiority between using graphical or textual nota-
tion for describing business process.
5 CONCLUSION
This work proposed a round-trip framework capa-
ble of automatically generating natural language text
from process models and updating these models from
text editions. It solves the problem of maintaining
text and model elements representation synchronized,
saves time and effort of analysts to manually write
text descriptions, enables domain experts to edit for-
mal process models without the efforts of learning a
business process modeling language.
The round-trip framework combines existing tools
from graph decomposition, natural language process-
ing and generation in an innovative way. The frame-
work provides the foundations for integrating textual
and model-based information, meeting the challenge
of processing both textual process descriptions and
models.
From a practical perspective, the framework helps
organizations to simplify business process manage-
ment since: (i) Creates automatically business pro-
cess instructions in natural language text; (ii) Allows
update the process model from text editions; (iii) Re-
duces efforts required by a business analyst which can
use the text to understand the processes.
The first experiment’s concluded the textual work
instructions can be considered equivalent, in terms of
knowledge representation, to process models within
an acceptable threshold. 74% of the subjects claims
the equivalence between both knowledge representa-
tions vary from 100% to 68%, and 86% of the sub-
jects claims the textual descriptions vary from ex-
cellent to good. This result is aligned with what
we expected due to the use of NLG techniques like
Discourse Marker insertion and Referring expression
generation, which are capable of enhancing the text
and improving its readability. The second experiment
concluded the knowledge represented by the manu-
ally updated text can be considered equivalent to the
automatically updated process model after the syn-
chronization within an acceptable threshold. 78%
of the subjects claims knowledge represented by the
manually updated text is equivalent to the automati-
cally updated process model.
Despite these encouraging results, the framework
is able to read process descriptions consisting of full
sentences. Another prerequisite is the text be gram-
matically correct, does not contain questions, and
has little process irrelevant information. Also, Word
Sense Disambiguation and Named Entity Recognition
was not within the scope. Another issue is the test
data set, which comprised thirty text-models. It is a
relatively small dataset. Therefore, this research’s test
results are not fully generalizable.
Some future work are: evolve the framework to
a “human in the loop” approach through the interac-
tion with domain experts to provide missing informa-
tion interactively, and further relive the business ana-
lyst from interviewing tasks; support specification or
overriding the text patterns; analyze and evolve it to
process large amounts of textual information, cover
other BPMN elements, include charts and other data
models, and add new languages.
ICEIS 2018 - 20th International Conference on Enterprise Information Systems
490
REFERENCES
Curtis, B., Sheppard, S. B., Kruesi-Bailey, E., Bailey, J.,
and Boehm-Davis, D. A. (1989). Experimental eval-
uation of software documentation formats. Journal of
Systems and Software, 9(2):167–207.
de AR Goncalves, J. C., Santoro, F. M., and Baiao, F. A.
(2009). Business process mining from group stories.
In Computer Supported Cooperative Work in Design,
2009. CSCWD 2009. 13th International Conference
on, pages 161–166. IEEE.
Ehud Reiter, R. D. (1997). Building applied natural lan-
guage generation systems. Natural Language Engi-
neering 1.
Friedrich, F., Mendling, J., and Puhlmann, F. (2011). Pro-
cess model generation from natural language text.
In Advanced Information Systems Engineering, pages
482–496. Springer.
Ghose, A., Koliadis, G., and Chueng, A. (2007). Rapid
business process discovery (r-bpd). In Conceptual
Modeling-ER 2007, pages 391–406. Springer.
Haisjackl, C. and Zugal, S. (2014). Investigating differ-
ences between graphical and textual declarative pro-
cess models. In Advanced Information Systems Engi-
neering Workshops, pages 194–206. Springer.
Kop, C., V
¨
ohringer, J., H
¨
olbling, M., Horn, T., Mayr, H. C.,
and Irrasch, C. (2005). Tool supported extraction of
behavior models. In ISTA, volume 63, pages 114–123.
Leopold, H., Eid-Sabbagh, R.-H., Mendling, J., Azevedo,
L. G., and Bai
˜
ao, F. A. (2013). Detection of naming
convention violations in process models for different
languages. Decision Support Systems, 56:310–325.
Leopold, H., Mendling, J., and Polyvyanyy, A. (2012).
Generating natural language texts from business pro-
cess models. In Advanced Information Systems Engi-
neering, pages 64–79. Springer.
Leopold, H., Mendling, J., and Polyvyanyy, A. (2014). Sup-
porting process model validation through natural lan-
guage generation.
Mel’
ˇ
cuk, I. A. and Polguere, A. (1987). A formal lexicon
in the meaning-text theory:(or how to do lexica with
words). Computational linguistics, 13(3-4):261–275.
Mendling, J., Reijers, H. A., and Cardoso, J. (2007). What
makes process models understandable? In Business
Process Management, pages 48–63. Springer.
Mendling, J., Reijers, H. A., and Recker, J. (2010). Activity
labeling in process modeling: Empirical insights and
recommendations. Information Systems, 35(4):467–
482.
Ottensooser, A., Fekete, A., Reijers, H. A., Mendling, J.,
and Menictas, C. (2012). Making sense of business
process descriptions: An experimental comparison of
graphical and textual notations. Journal of Systems
and Software, 85(3):596–606.
Pree, W. (1994). Meta patterns a means for capturing the es-
sentials of reusable object-oriented design. In Object-
oriented programming, pages 150–162. Springer.
Rodrigues, R. D. A., Barros, M. D. O., Revoredo, K.,
Azevedo, L. G., and Leopold, H. (2015). An exper-
iment on process model understandability using tex-
tual work instructions and bpmn models. In Software
Engineering (SBES), 2015 29th Brazilian Symposium
on, pages 41–50.
Sinha, A. and Paradkar, A. (2010). Use cases to pro-
cess specifications in business process modeling no-
tation. In Web Services (ICWS), 2010 IEEE Interna-
tional Conference on, pages 473–480. IEEE.
van der Molen, T. (2011). Maintaining consistency between
business process diagrams and textual documentation
using the epsilon model management platform.
Vanhatalo, J., V
¨
olzer, H., and Koehler, J. (2009). The re-
fined process structure tree. Data & Knowledge Engi-
neering, 68(9):793–818.
Wang, H. J., Zhao, J. L., and Zhang, L.-J. (2009). Policy-
driven process mapping (pdpm): Discovering process
models from business policies. Decision Support Sys-
tems, 48(1):267–281.
Zugal, S., Pinggera, J., and Weber, B. (2011). Assessing
process models with cognitive psychology. In EMISA,
volume 190, pages 177–182.
BPMN Model and Text Instructions Automatic Synchronization
491