Você está na página 1de 17

Requirements Eng

DOI 10.1007/s00766-012-0155-2

ORIGINAL ARTICLE

Comparing the impact of the OO-DFD and the Use Case methods
for modeling functional requirements on comprehension
and quality of models: a controlled experiment
Michal Dahan Peretz Shoval Arnon Sturm

Received: 2 April 2011 / Accepted: 4 May 2012


 Springer-Verlag London Limited 2012

Abstract Users requirements of an information system improved method sketch that we call Enhanced Use Case,
are modeled in the analysis phase of the development which will be evaluated in future work.
process. The requirements can be modeled with various
modeling methods. In this study, we compare two alter- Keywords Functional analysis  IS development 
native methods for modeling the functional requirements: Modeling methods  User requirements  Use Case  UML 
one is the UML Use Case (UC) model; the other is OO- FOOM  OO-DFD
DFD transaction (Object-Oriented DFD is a variant of DFD
that includes data classes rather than traditional data
stores). Each of these modeling methods consists of dia- 1 Introduction
grams accompanied with narrative, semi-structured
descriptions explaining their details. We conducted a Elicitation and modeling the users requirements from an
controlled experiment that compared the comprehension of information system is a very important task, for which
the two models (i.e., the diagrams and their descriptions) of many methods were introduced over the years. A few
a certain system and the quality of models created for a examples are discussed below.
certain system with each of the two modeling methods. The i* [37] is a framework that attempts to deal with activities
main results of the experiment are that models created with that precede the formulation of the initial requirements.
the UC method are of better quality than models created These early phase requirements engineering activities
with the OO-DFD transaction method because the former include those that consider how the intended system would
are simpler and less detailed; creating highly detailed meet organizational goals, why the system is needed, what
models are error prone. Interestingly, in spite of the dif- alternatives might exist, what the implications of the alter-
ference in the level of detail and structure, the experiment natives are for various stakeholders, and how the stake-
reveals no significant difference in comprehension of holders interests and concerns might be addressed. The
models of the two methods. The results call for improve- emphasis in i* is on understanding the whys that underlie
ment of the modeling methods in a way that considers the system requirements, rather than on the precise and detailed
advantages of each of them, and thus we propose an specification of what the system should do, which is
modeled with other modeling methods.
Business Process Modeling Notation (BPMN) [35]
provides a notation that is readily understandable by all
M. Dahan  P. Shoval (&)  A. Sturm
business users, from business analysts who create the initial
Department of Information Systems Engineering,
Ben-Gurion University, 84105 Beer-Sheva, Israel drafts of the processes, to technical developers who are
e-mail: shoval@bgu.ac.il responsible for implementing the technology that will
M. Dahan perform those processes, and finally, to the business people
e-mail: dahanmic@bgu.ac.il who will manage and monitor those processes. BPMN
A. Sturm defines a Business Process Diagram (BPD), which is based
e-mail: sturm@bgu.ac.il on a flowcharting technique tailored for creating graphical

123
Requirements Eng

models of business process operations. Business process oriented methodologies. Some of the famous ones were
modeling is used to communicate a wide variety of infor- OOA/OOD [7], OMT [29], and Booch method [4]. Rumb-
mation to different audiences and provides much more aughs and Boochs methodologies became very popular,
information on the business process than just modeling the and in 1995, they joined forces and created the Unified
functional requirements of the users from the system. Method (UM). Later on, Jacobson joined the two with his
The early systems analysis and design methodologies, comprehensive method ObjectOry [17], thus creating the
which emerged in the late 1970s, utilized Data Flow Dia- Unified Modeling Language (UML) [5]. Since the mid
grams (DFD) as their main modeling technique. Most 1990s, UML became a de facto standard for modeling
notable examples for such methodologies are DeMarco systems based on the OO approach. UML consists of over 15
[10] and Gane and Sarson [12]. One of the biggest prob- types of diagrammatic modeling methods. One of the most
lems with the traditional methodologies was in the popular of them, which is used for modeling the users
transition from the analysis phase to the design phase, functional requirements, is the Use Case (UC) model.
which was not smooth and caused many problems. Spe- In this study, we compare the UC model of UML and
cifically, it was not easy to convert the DFDs, products of the OO-DFD transaction diagrams of FOOM. The two
the analysis phase, to Structure Charts, which were then a models have many things in common. In essence, a UC and
common method for designing the application programs. In a transaction have the same objective: to define and
order to deal with this and other problems, Shoval [30] describe an independent task performed by the system to
developed the ADISSA methodology, which enables a support a user. The similarity between the two modeling
smooth transition from the analysis phase and its DFDs to methods was also noted by Ramsin and Paige [27] who
the design phase. This was done by defining the concept of compared various object-oriented development methodol-
DFD transactions. A DFD transaction is an independent ogies. Yet, the two methods differ in the variety of com-
process performed by the information system to support a ponents included in the respective diagrams, and in the
users task. According to ADISSA, all transactions of a structure of the description accompanying the diagrams (as
system are derived from the DFDs: each transaction con- will be elaborated later on). The main objective of this
sists of one or more elementary functions that are chained study is to compare between these alternative functional
by data flows, and of external entities and data stores that requirements modeling methods and to find out and ana-
are too connected with data flows to these functions. The lyze the strengths and weaknesses of each.
process logic of each transaction is defined and described The importance of comparing alternative modeling
using pseudo-code. Later on in the design phase, the methods stems also from the evolution of the model-driven
transaction descriptions are more detailed and other com- development approach, which claims that models (rather
ponents of the system are designed, notably the application than the code) serve as the systems backbone. Accord-
programs, the user interface (menus trees), the input and ingly, it is important to compare and analyze the differ-
output screens and reports, and the relational database ences between competing modeling methods. Modeling
schema. methods can be similar or different from various perspec-
With the emergence of Object-Oriented (OO) develop- tives; in this study we compare between the two modeling
ment methodologies, Shoval enhanced ADISSA and created methods mainly from the point of view of understanding
FOOMFunctional and Object-Oriented Methodology, models and creating correct models. Such comparisons are
which combines the functional and OO approaches [31, 32]. important not only from an academic viewpoint but also
According to FOOM, two main models are created in the from the pragmatic one, because it might guide analysts/
analysis phase: a conceptual data model and a functional designers regarding which modeling method to adopt and
model. The conceptual data model is an initial class diagram use. Furthermore, due to insights gained regarding the
that consists of the data classes, attributes, and various strengths and weaknesses of the two modeling methods, we
relationships between classes (but with no methods yet). The are able to propose how to improve each of them and also
functional model consists of OO-DFDs, which are similar to to propose a hybrid modeling method that combines the
the traditional DFDs, but instead of data stores, they include strong parts of each of them.
data classes, which are taken from the initial class diagram. The rest of this paper is structured as follows: Sect. 2
As in ADISSA, the OO-DFDs are then decomposed into provides a background on the two modeling methods and
transaction diagrams, and then their process logic is descri- surveys related studies on empirical evaluations of mod-
bed. Later on, the transaction descriptions are decomposed eling methods, specifically on comparisons between the
into class methods, which are attached to proper classes in DFD and UC methods. Section 3 describes the planning of
the class diagram. the controlled experiment, and Sect. 4 presents the results.
In the early 1990s, many OO analysis and design meth- Section 5 discusses the evaluation of the results, presents a
odologies emerged, replacing the traditional, functional- new proposal for a combined modeling method, and

123
Requirements Eng

discusses threats to the validity of the experiment. Finally, sometimes, there is a UC bubble that is connected to other
in Sect. 6, we discuss the experiment conclusions and set UC bubbles. The main types of connections are include
the plans for future research. and extend. (More details about these types of UCs are
beyond the scope of this paper).
Many ways have been published over the years to
2 Background describe UCs, for example, [8, 9, 21]. Figure 2 shows a UC
diagram that is assumed to portray the same transaction as
2.1 Two functional modeling methods: OO-DFD of Fig. 1, while Table 2 shows the description of this UC.
and Use Case
2.1.3 Differences between the two methods
2.1.1 OO-DFD
As said, the above two modeling methods have the same
An OO-DFD, according to FOOM, is a DFD adapted to the purpose: to describe the functional requirements of an
OO world: instead of data stores, it includes data classes information system. As can be seen in the above examples,
classes that are already defined in the initial class diagram the methods differ both in the diagrams and in the
of the application, which is created prior to that. Then, descriptions. The major difference is in the details included
using certain rules, the OO-DFDs are decomposed into in each diagram. A process in the UC diagram is just a
transaction diagrams. Each transaction consists of one or bubble (as mentioned, sometimes a few bubbles connected
more elementary functions that are chained by data flows, to each other with include/extends connections) that is
and of the external entities and data classes that are too connected to the role (actor) or roles that interact with the
connected with data flows to those functions. After this system in order to operate the process. Contrarily, an
decomposition, the process logic of each transaction is OO-DFD transaction diagram usually consists of a series of
defined and described. At first, a top-level (i.e., general) functions that are chained to each other with directed data
description is created. Later on, the transaction description flows and of data classes and user/external entities that are
is more detailed, taking into consideration the inputs and connected to those functions with data flows. The data
outputs/reports involved in each transaction; and the flows carry data elements from a source component to a
reading and writing of data from/to the data classes. In a destination component. Note that the data classes included
later stage, each transaction description is decomposed into in the OO-DFD transaction diagrams are taken from the
methods that are attached to proper classes. But this paper initial class diagram. Another difference relates to the
deals only with the functional modeling aspect of this external entities: in an OO-DFD transaction, the external
methodology, as expressed by transaction diagrams and entities, also termed user entities, are the sources of input
their top-level descriptions. data or the destinations of information produced by the
Figure 1 shows an example of a transaction diagram that transactionbut they are not necessarily the actual oper-
has been extracted from an OO-DFD that models the IFIP ators of the transaction. Contrarily, actors in the UC dia-
conference system [22], while Table 1 shows its top-level grams signify the operators of the UC who interact with it.
description. Note that the description refers to the various The difference between the descriptions follows from the
components included in the transaction diagram. differences between the diagrams: a UC description
includes many details that are not included in the diagram;
2.1.2 Use Case this is actually the only place where the meaning of the
UC is defined. Contrarily, the OO-DFD transaction
The Use Case method (UC) was introduced by Jacobson description repeats details that can be seen in the diagram,
[16] and then became part of UML. Nowadays, UC is one that is, there is some redundancy. Another difference is in
of the most popular techniques to describe the functional the description of the process logic: in the UC description
requirements of a system in the OO world. The main (where it appears under the Main success scenario and
components of a UC diagram are the UC bubbles and the Extensions compartments), the process logic is written
actors. A UC bubble is a process that executes from an in natural-language-like fashion, while in the OO-DFD
external point of view, while a UC actor represents a role transaction, the process logic (appearing under the Top-
of an entity that interacts with the system, that is, that level description compartment) is described using more
operates the UC. A common way to define UCs is to start pseudo-code-like style, referring to the various components
from identifying the actors of the system and then think of of the OO-DFD transaction and using standard structured
all ways those actors will use the system [1]. Actors are programming patterns (not shown in the above example).
connected with UC bubbles that they interact with, but the In essence, the UC diagram includes very little informa-
connections are (usually) undirected. In a UC diagram tion, while the UC description includes the details; on the

123
Requirements Eng

Details of reviewers and non


1
assigned articles
Display details of U1
reviewers and non PC Chair
Non
assigned articles Reviewers
assigned
articles details

C1 - Article C3 - Reviewer C6 - Article review

Details of
Reviewers and non
assigned articles Assignment
Status=
sent for review details

Articles details and last date


Chosen article and assigned
2 to submit review report
U1 reviewers U2
Assign reviewers to
PC Chair Reviewer
a chosen article

Fig. 1 Transaction diagram extracted from an OO-DFD

Table 1 Top-level description of the transaction


Component Description
Assign reviewers to
a submitted article
Transaction Assign reviewers to a submitted article
name PC Chair Electronic mail system
Transaction User transaction {user transaction means an
type interactive transaction, in which a use/operator Fig. 2 UC diagram
interacts with the system}
Top-level Begin transaction 2.2 Related studies
description Read from class C1 (Article): non-assigned articles
Read from class C3 (Reviewers): reviewers details 2.2.1 Empirical evaluations of modeling methods
Execute function 1display details of reviewers
and non-assigned articles Over the years, many studies have been published on
Output to U1 (PC Chair): details of reviewers and comparisons between different modeling methods,
non-assigned articles
Move to function 2: details of reviewers and non- Table 2 Description of the UC
assigned articles
Component Description
Input from user U1 (PC Chair): chosen article and
assigned reviewers UC name Assign reviewers to a submitted article
Execute function 2assign reviewers to a chosen Description The PC Chair assigns reviewers to a submitted but not
article yet assigned article
Write to class C1 (Article): Status = sent for Actors PC Chair, electronic mail system
review Preconditions There are submitted articles in the system that have not
Write to class C6 (Article review): assignment yet been assigned to reviewers
details Postconditions The chosen article is assigned to reviewers
Output to U2 (Reviewer): articles details and last Main success 1. The system displays details of reviewers and non-
date to submit review report scenario assigned articles to the PC Chair
End transaction 2. The PC Chair selects and inputs to the system an
article from the list and the reviewers assigned to
review it
3. The assignment details are saved in the system
other hand, an OO-DFD transaction diagram includes 4. The system sends the articles details and the last
many details, while the transaction description includes date to submit review reports to the reviewers, by the
electronic mail system
redundant information that is presented in a structured
Extensions None
format.

123
Requirements Eng

including comparisons between different methods and diagrams than OMT analysts. Kabeli and Shoval [19]
frameworks for modeling users requirements. compared OPM and FOOM in two controlled experiments
Bass et al. [3] evaluated five methods for elicitation and from the point of view of users and analysts. Their main
expression of requirements with respect to their ability to results were that FOOM analysis specifications are more
capture architecturally significant requirements. The eval- comprehensible to users than those based on OPM,
uated methods included: natural language, use case, quality whereas in terms of the data model, there is no significant
attribute workshop (QAW), global analysis, and OBriens difference. From the analysts point of view, they com-
approach. The main conclusions were that QAW is the pared quality, that is, correctness of specifications. The
most expressive method for specifying unambiguous main results of these experiments were that analysts
requirements, and that OBriens approach is the only create more correct specifications with FOOM method-
method that addresses business goals explicitly. The other ology than with OPM.
methods rely on the stakeholders to translate business and
mission goals into requirements. 2.2.2 Empirical comparisons of DFD and UC modeling
Hadar et al. [14] present an empirical study that com- methods
pares Use Case and Tropos, a goal-oriented approach. The
authors objective was to evaluate different levels of As we have seen, many studies have evaluated and com-
comprehension of requirements models expressed in both pared different modeling methods that may be used for the
methods, as well as to estimate the time required to per- same purpose. Therefore, it is only natural to compare
form simple analysis tasks. Preliminary results show that the DFD with the UC modeling methods, since they have
Tropos models seem to be more comprehensible to novice the same purpose: to define the users functional require-
requirements analysts, although more time consuming than ments. In this section, we review studies that have com-
Use Case models. pared between these two methods.
Topi and Ramesh [34] surveyed many studies pub- Millet and Nelson [23] compared DFDs and UCs using
lished between 1978 and 2001 that employed laboratory questionnaires. During 20032006, the course of system
experiments to evaluate the usability of data models/ analysis was offered 12 times, each time to different stu-
methods. They found out that the most frequent inde- dents but all times by the same lecturers. In each semester,
pendent variable in those studies was the data model; the the course was given to two different groups in parallel,
next category of independent variables consists of user and the researches split the students into two groups: stu-
characteristics, for example, experience, education, and dents in one group had to perform a DFD task and then a
intellectual ability; other independent variables are task UC task, and students in the other group had to perform a
characteristics, for example, comprehension and task UC task and then a DFD task. All tasks included only
complexity. The dependent variables are mostly model diagrams, without descriptions. The first task of the first
correctness, time used to create the model, declarative group was to analyze with several DFDs a simple work
knowledge (understanding of the notation), and attitude. order system including only 3 processes and 3 external
In most cases, the correctness of a model has been entities, while using a certain CASE tool; the participants
measured according to the degree to which it corresponds had 2 days to complete this task. Afterward, the partici-
to a predefined solution. Attitude includes mainly pref- pants learned the UC method and were given the second
erence to use a certain model and perceived ease of use. taskto analyze the same system with UCs using another
There are studies that take an analyst/designer/modeler CASE tool; the participant had 2 days to complete this
perspective; these are mainly concerned with measuring task. For the second group, the order of learning and tasks
performanceusually model correctness. Other studies was the opposite. After completing the two tasks, each
that take a user perspective are mainly concerned with participant was asked to answer a questionnaire with five
measuring comprehensibility of models and preference of claims on each of the methods, using a 17 point scale. For
models by users. For example, Peleg and Dori [26] example, one claim was: This methodology is easy to
compared two methodologies: OPM/T, a variant of OPM understand. The results showed that the participants found
for real-time systems, and OMT/T, a similar variant of the DFD method better in its effect on helping systems
OMT. The study included comparison of both data analysts communicate requirements to programmers, but
modeling and process specifications from the points of they did not find differences between the methods in all
view of comprehension and quality. The authors found other claims.
that in more cases, OPM specifications were more com- This research has several drawbacks. One is the use of a
prehensible than OMT specifications. The authors also questionnaire only, that is, the results are based solely on
found that OPM analysts produced significantly better the participants opinions regarding the five claims.

123
Requirements Eng

Although the tasks of the participants were to analyze object-oriented development process. Note that we do not
(though very small) systems, the correctness or quality of compare UCs with OO-DFDs, but rather with the trans-
their solutions were neither examined nor compared. actions that are derivable from them, because actually it is
Another problem is that the students had to use two dif- the transactions that are equivalent to UCs, while DFDs
ferent CASE tools; there is a possibility that the learning have other components and characteristics that do not
process and usability of these tools had some impact on the exist in UCs (for example, DFDs express hierarchical
students opinions about the modeling methods. Note also decomposition of functionality, which has no equivalent
that the students had 2 days to complete each task; it means in UCs). In the next sections, we describe the experiment
that there was not enough control on the process of solving planning and analyze the results, following the guidelines
the tasks by the students. Finally, this experiment dealt for reporting experimental research as specified by
only with comprehension of diagrams. Wohlin et al. [36].
Jeyayaj and Sauter [18] too compared between the DFD
and the UC modeling methods. In their experiment, the
participants were 4 classes of business administration stu- 3 Experiment planning
dents, who did not learn the examined methods, and 5
classes of MIS students, who did learn and practiced with 3.1 Goals
the examined methods. The business administration stu-
dents were considered novice users, while the MIS students As said, our objective is to compare the two alternative
were considered experienced users. The differences functional modeling methods: UCs and their descriptions
between the methods were examined from the point of versus OO-DFD transaction and their descriptions. The
view of comprehension of diagrams. A student registration comparison between the methods is performed based on six
system was modeled into a DFD diagram and a UC dia- goals:
gram. (Each diagram included several DFDs/UCs, but there
G1: Comprehension of modelswe ask whether there is
is no mention of their descriptions). Each participant was
a difference in the understanding of the models (i.e.,
asked to explain in free text what he/she understands from
diagrams and descriptions) by users.
each diagram. The participants were given a week to
G2: Quality of created modelswe ask whether there is
complete the first task and then another week to complete
a difference in the quality of models created by
the second task. The participants (of either type) were split
modelers.
into two groups, one receiving the DFD task prior to the
G3: Time taken to understandwe ask whether there is
UC task and the other receiving the UC task prior to the
a difference in the time it takes to understand given
DFD task. The overall conclusion of this research was that
models.
for experienced users, the DFD method provides a better
G4: Time taken to modelwe ask whether there is a
tool to describe functional requirements, while for the
difference in the time it takes to create models.
novice users, no difference was found between the
G5: Perceived comprehension of modeling methods
methods.
we ask whether there is a difference in the comprehen-
This research too has several drawbacks. First, the par-
sion of the modeling methods as perceived by users.
ticipants were asked to provide free-text explanations of
G6: Perceived quality of modeling methodswe ask
what is seen in the given diagrams, which was then coded
whether there is a difference in the quality of the
into some numbers; this is not a precise way to measure
modeling methods as perceived by modelers.
comprehension. Secondly, the students had a week to
complete each task, so here too there was not enough We compared the two methods in a controlled experi-
control on the process. Third, this experiment too dealt ment in quasi-laboratory setting, where we address the
only with comprehension of diagrams. above goals. The experiment involved two tasks: the first
In summary, the above studies on the comparison task was of understanding diagrams and descriptions rep-
between DFDs and UCs suffer from many limitations. In resenting the functional requirements of a certain system;
particular, they deal only with comprehension of dia- the second task was of creating diagrams and descriptions
grams. In our research, we compare the two alternative from a requirements document of a certain (other) system.
modeling methods from the points of view of both The participants were divided into two homogeneous
comprehension and quality of model created by analysts. groups; each participant in each group had to perform the
In our comparison, we use OO-DFDs, which include data two tasks using one of the two methods: the first task was
classes rather than data stores in traditional DFDs; this of understanding given models, and the second, of creating
enables us to compare two methods that may be used models. Thus, we designed an experiment of one factor
equivalently for defining functional requirements in an with two treatments.

123
Requirements Eng

3.2 Hypotheses and the research models connections between them. Cognitive theories regarding
the comprehensibility and quality of diagrams can help
3.2.1 Comprehension of models assume which kinds of diagrams are more or less com-
prehensive. Several studies were done on comprehension
The research model for the comprehension task is pre- and quality of diagrams in conceptual modeling based on
sented in Fig. 3. cognitive theories, for example, [15, 20, 28]. One important
Following the research goals, the conjectures for this framework based on cognitive theory is COGEVAL pre-
experiment are as follows: sented by Rockwell and Bajaj [28]. The framework states
that a model that has less relationship information is less
1. There are no differences between the two methods
comprehensive than a model that has more relationship
regarding the level of comprehension (i.e., the under-
information. In our case, a UC diagram has less relation-
standing) of the diagrams and their descriptions.
ship information compared to an OO-DFD transaction
2. The UC method is superior to the OO-DFD transaction
diagram. An experimental study for evaluating the influ-
method regarding the time it takes to understand the
ence of the number of concepts (NOC) in a model on the
diagrams and their descriptions.
readability of the model [2] also supports this assumption;
3. The UC method is superior to the OO-DFD transaction
the experiments results showed that a higher NOC will
method regarding the perceived comprehensibility of
lead to a higher percentage of questions about the domain
the model.
that the subject can answer (we refer to this as higher
Theoretic rationale for the conjectures: According to a comprehension). Moreover, one of the guidelines for cre-
framework suggested by Gemino and Wand [13] for ating good modeling diagrams is that diagrams should not
evaluating modeling techniques, both grammar-based and exceed perceptual and cognitive limits in the sense that it
cognitive-based approaches should be considered comple- should not show too much information on a single diagram,
mentary in order to evaluate two (or more) different which may result in absurdly complex diagrams that are
modeling techniques. Grammar-based approaches identify a barrier rather than an aid to communication [24]. In our
the differences and generate predictions regarding gram- context, the OO-DFD transaction diagram contains a lot of
mar efficacy, while cognitive-based approaches suggest information that can make it more complex for under-
ways to observe the effects of grammar differences and test standing than the UC diagram. This is also supported by the
the predictions. We follow this framework and rationalize COGEVAL framework according to which a model with
our hypotheses by emphasizing the differences between the more elements will be more complicated to understand
two methods and then considering cognitive researches in than a model with fewer elements, due to the limited
order to predict the effect of a grammars expressiveness capacity of the short-term memory. An OO-DFD transac-
(the differences between the methods) on the effectiveness tion diagram has more elements than a UC diagram; hence,
of its use by individuals (i.e., understanding the model due to this proposition, it is expected to be less understood.
presented to the participants). The grammar- and cognitive- These arguments are valid also for the descriptions of
based rationales for our hypotheses are presented below. the diagram: an OO-DFD transaction description is full of
An OO-DFD transaction diagram contains many details; details; it mentions the various components of the OO-DFD
it usually includes several functions, data classes, external transaction and specifies the process logic using pseudo-
entities, and directed data flows between them. In contrast, code. In contrast, a UC description is less structuredmore
a UC diagram contains only UC bubbles, actors, and similar to natural language. So, on the one hand, an
OO-DFD transaction description includes more informa-
tion than a UC description, but on the other hand, it might
be complicated to process. Hence, the dilemma is which
of the conflicting forces is stronger: simplicity but less
information versus complexity but more information? more
natural description or more structured description?
Recall that in the above two surveyed studies that dealt
with comprehension, the researchers concluded that DFDs
are to a certain degree superior to UCs. We speculate that
due to the above contrasts and based on cognitive theories,
there will be no difference between the two modeling
methods in understanding the diagrams and their descrip-
tions. Our speculation is further supported by Gemino and
Fig. 3 The comprehension research model Wand [13], who claim that while one modeling grammar

123
Requirements Eng

Table 3 The comprehension hypotheses


Aspect H0 H1

Comprehension of models H01: Comp (UC) = Comp (DFD) H11: Comp (UC) = Comp (DFD)
Time taken to understand models H03: TComp (UC) = TComp (DFD) H13: TComp (UC) = TComp (DFD)
Perceived comprehension of modeling methods H05: PComp (UC) = PComp (DFD) H15: PComp (UC) = PComp (DFD)

may be highly expressive and hence superior in a grammar- 3. The UC method is superior to the OO-DFD transaction
based evaluation (in our case, OO-DFD transactions are method regarding the perceived quality of the created
grammatically superior to use cases), a representation model.
created with that grammar might be overly complicated,
Theoretic rationale for the conjectures: Based on the
leading to difficulties in developing domain understanding
grammatical differences between the two methods and on the
by the person viewing the model.
cognitive theories mentioned above, the COGEVAL
As for the time it might take to understand the models,
framework states that when a model requires a greater
we believe that due to the simplicity of the UC diagram and
number of simultaneous items to create the diagram, the
the more natural language used in the description, it will
diagram is of less quality than a diagram of a model that
take less time to understand UC models. Bajaj [2] also
requires fewer numbers of items. An OO-DFD transaction
shows that a higher NOC will lead to more time to answer
diagram includes more types of components compared to a
questions regarding schemas. This finding supports our
UC diagram, and so does the OO-DFD transaction descrip-
speculation that it will take less time to answer compre-
tion compared to a UC description. Moreover, a UC
hension questions in the UC model. For the same reasons,
description is simpler to create because it is mostly in natural
we speculate that users will perceive the UC method as
language, less detailed and less structured, while an
better than the alternative.
OO-DFD transaction description includes all the details/
To examine the various conjectures, we formulate the
components of the respective diagram, and the process logic
hypotheses as appears in Table 3.
is expressed in pseudo-code. This implies that with UCs,
there are fewer chances to commit errors while writing the
3.2.2 Creation of models
descriptions. Hence, we speculate that the diagrams and
descriptions created with the UC method will be of better
The research model for the model creation task is presented
quality than the alternative. Our speculation is also supported
in Fig. 4.
by Bajaj [2], who claimed that increasing the number of
Following the research goals, the conjectures for this
concepts (NOC) in a model makes it harder to create model
experiment are as follows:
schemas from an analyst point of view. Hence, OO-DFD
1. The UC method is superior to the OO-DFD transaction transaction diagrams and description, which have more
method regarding the quality of the diagrams and concepts than UC diagrams and descriptions, will be harder
descriptions created by analysts. to create and thus will have more errors. For the same reason,
2. The UC method is superior to the OO-DFD transaction we assume that it will take less time to create UC models and
method regarding the time it takes to create the that analysts will perceive the UC method as better for
diagrams and descriptions. defining user requirements.
To examine the various conjectures, we formulate the
hypotheses as appears in Table 4.

3.3 The independent variables

The independent variable is the modeling method: UC or


OO-DFD transaction.

3.4 The dependent variables

a. In the comprehension experiment, the dependent


variables are as follows:
1. Comprehension of diagrams and their descriptions.
Fig. 4 The quality of model creation research model This variable is measured by a comprehension

123
Requirements Eng

Table 4 The model creation hypotheses


Aspect H0 H1

Quality of models H02: Create (UC) = Create (DFD) H12: Create (UC) = Create (DFD)
Time taken to create the models H04: TCreate (UC) = TCreate (DFD) H14: TCreate (UC) = TCreate (DFD)
Perceived quality of models H06: PCreate (UC) = PCreate (DFD) H16: PCreate (UC) = PCreate (DFD)

questionnaire consisting of true/false/cant tell state- descriptions of a certain information system; in the model
ments about information included in the diagrams and creation task, all participants received the same require-
their descriptions. The same questions were used to ments document of a certain (other) system and had to
measure the comprehension of the two models. create the models using one of the two modeling methods
2. Time taken to comprehend. This variable is measured (the same method that they used in the comprehension
by the time (in minutes) taken by participants to task).
complete the comprehension task.
3. Perceived comprehensibility of models, that is, the 3.5.1 The subjects
participants opinion regarding the ease of understand-
ing the diagrams and descriptions. This variable is Fifty-three students participated in the experiment that took
measured using a 7-point scale. the form of a bonus test at the end of the semester, in
b. In the modeling experiment, the dependent variables which they have studied the two modeling methods in two
are as follows: mandatory courses that were given in parallel: in the
1. Quality of the diagrams and their descriptions. The Information Systems Analysis & Design course, they
participants create models based on a requirements studied OO-DFDs and transactions as part of the FOOM
document of a certain (other) system. The quality of a methodology; in the Object-Oriented Analysis & Design
created model is the degree to which the diagrams and course, they studied the UC method as part of UML. Each
descriptions describe correctly and accurately the method was taught by an experienced lecturer.
requirements of the system. Quality is determined by Before the experiment began, the subjects were ran-
graders who were assisted with an expert solution. domly divided into two treatment groups as shown in
2. Time taken to create models. This variable is measured Table 5. In these kinds of empirical experiments, there are
by the time (in minutes) taken by participants to many factors that may influence the results, such as the
complete the model creation task. personal characteristics of the subjects and their experi-
3. Perceived quality of modeling methods, that is, the ence. As in most experiments of this form, neutralizing
participants opinion regarding how good is the such factors is done by randomization. Note also that third-
modeling method for requirements specification. This year students have almost no experience, so this variable
variable too is measured using a 7-point scale. could have no effect in this study.
As said, subjects in each group were assigned to work
with one modeling method: subjects in group G1 per-
3.5 The controlled variables
formed the two tasks using the OO-DFD transaction
method, while subjects in group G2 performed the two
The controlled variables in the two experiments are the
tasks using the UC method. The subjects were motivated
participants and the tasks. The participants are a homogenous
by adding bonus points to their final grades in the course in
group of third-year students of software engineering who
which they studied the method that they were assigned to in
studied the same courses. They were split randomly into the
the experiment. The number of bonus points given to each
two treatment groups, where members in each group used
student was determined according to his/her performance
only one of the two modeling methods. This way we avoided
in the two tasks. (The procedure was approved by the
possible biases that might have affected this variable (e.g.,
Ethics committee).
age, sex, and experience). Note that this kind of assignment
of participants to treatment groups is often used in experi-
mental evaluation of modeling techniques. Table 5 Assignment of subject into treatment groups
The control over the tasks variable is achieved by giving Group Size Task 1: comprehension of Task 2: model
the participants in the two groups the same tasks; only the models creation
modeling methods they used were different: in the com-
G1 28 OO-DFD transaction diagrams and descriptions
prehension task, all participants had to answer exactly the
G2 25 Use Case diagrams and descriptions
same questionnaire while given different diagrams and

123
Requirements Eng

3.5.2 The tasks Table 6 Examples of statements in the questionnaire


When a clients order is updated, first the updated order details are
In the first taskcomprehension of modelseach subject inserted and then the card details are read
received diagrams and their respective descriptions of an The client does not need to approve the chosen design details
information system, according to his/her group. The system before ordering the desired amount
for this task was Greeting Cards Ordering (similar to the A message is sent to the client if his order is not updated according
one used in Dori and Goodman [11] and in Kabeli and to his request
Shoval [19]). The domain of this system was not familiar to When a client wants to choose a card from the existing variety, the
system will display to the client all the cards existing in the
the students (unlike, for example, a university management
collection
system would), and thus, there is no concern that the par-
ticipants will answer the questions due to background
knowledge of the domain [6]. This guideline is also men- model (diagrams and descriptions) he/she used. This
tioned in Parsons and Cole: subject matter experts should questionnaire consisted of two parts: (a) a 17 ordinal scale
not be used, it is critical that participants can answer question, where 1 means the model is very difficult to
questions by using only the script, rather than by using understand and 7 means the model is very easy to
background knowledge [25]. For the OO-DFD transaction understand; (b) an open-ended question, where the subject
method, we created a flat OO-DFD of the system consisting was asked to write comments about the comprehensibility
of 5 transactions, along with a description of each trans- of the modeling method he/she worked with.
action. Similarly, we created an equivalent UC model that In the second task of the experimentmodel creation
included respective UC diagram and descriptions. The each subject received a narrative requirements document of
diagram and descriptions were examined by experts of the an information system. The task was to create diagrams
two methods to verify that they are correct and in accord (OO-DFD transaction or UCs) and respective descriptions.
with the respective method.1 (Recall that each subject continued to work with the same
Along with the diagrams and descriptions, each partic- modeling method as in the first task so that he/she is not
ipant received a questionnaire consisting of 22 questions. confused with a different method). The requirements doc-
The same questionnaire was given to all subjects in the two ument was part of the IFIP Conference System (used also
groups. The questions were actually statements dealing in [19, 22]). Prior to the experiment, two detailed solutions,
with facts appearing in the diagrams and their descriptions. one for each method, were created by experts of each
For each statement, each subject had to mark whether it is method. These solutions were used after the experiment by
true, false, or cant tell. As already said, the the graders who graded the models that were created by the
methods are not information equivalent [6, 33]. For subjects.
example, data classes appear in transactions, but not in Along with the requirements document, each subject
UCs. Parsons and Cole claim: if one form provides received an instructions form that included explanations
enough information to answer selected questions correctly about what he/she is expected to do, and an example of a
while a second form does not, it would not be surprising to diagram (a transaction or a UC) and a respective descrip-
find that participants receiving the first form outperform tion. The example was provided in order to help the par-
those receiving the second form on those questions [25]. ticipants remember the notations and structures used in the
We agree with this claim and overcame the problem by modeling method.
asking questions that do not refer to components not Once completed this task, each participant was given
appearing in the two models. For example, instead of another post-test questionnaire, where he/she was asked to
referring to a certain data class used to store or retrieve express his/her opinion about how good is the method to
data, as might appear in a transaction only, we just refer to create a functional model of a system. Here too we used a
information items. Table 6 presents a few examples of 17 ordinal scale question (1very bad, 7very
statements included in the questionnaire. good) and an open question to write comments about the
Once completed the questionnaire, each subject was method.
given a post-test questionnaire where he/she was asked to It has to be noted that a few weeks prior to the actual
express his/her opinion about the comprehensibility of the experiment, we have conducted a pilot test with 4 senior
students who studied and practiced with the two methods.
1
In the Comprehension task we face a potential bias since the The pilot test helped us to verify the correctness of the
researchers created the models. This problem is well known since in tasks and the clarity of the questions, and to assess the time
many other experimental studies the researchers create the modeling
needed to complete the two tasks. Based on that, we have
tasks. Our work was done honestly and we created the best possible
models with each of the methods, with no bias (as can be proved by decided to allocate a total of 3 h for the two tasks. We told
the results of the comprehension task in next sections). the subjects in advance that they have 3 h to complete the

123
Requirements Eng

Table 7 The comprehension results


OO-DFD UC Statistical analysis results Conclusion
transactions (a = 0.05)

Comprehension grade 70.71 % 70.68 % p = 0.98 (t test) No difference in comprehension


Time to complete comprehension 43 min 35 min p = 0.008 (t test) Difference in favor of UC
task method
Perceived comprehensibility 5.18 4.28 p = 0.023 (Wilcoxon) Difference in favor of OO-DFD
of models transaction method

two tasks, but we advised them to dedicate about an hour to results showed high correlations: for the graders of the
the first task (comprehension) and the rest, to the second OO-DFD transaction, the Pearson correlation coeffi-
task (model creation). cient was 0.81; for the UCs, it was 0.76. Therefore, we
could average the two grades and declare it as the
3.6 The analysis procedure: grading the solutions quality grade of each solution.

The grades of the solutions of the two tasks were deter-


mined as follows:
4 Analysis
Comprehension task: The comprehension task included
a questionnaire containing 22 statements having one Prior to testing the differences between the results using
correct answer each (True or False). For each solution, parametric t tests, some assumptions had to be validated. The
the number of correct answers was counted. t test assumptions are that: (a) the samples are independent;
Model creation task: Grading the solutions of this task (b) the samples distribution is normal; (c) the homoscedas-
was more complicated, because for a given require- ticity assumption is valid. The independency assumption is
ments document, many correct solutions (i.e., func- validated due to taking the observations from an independent
tional models) are possible. This requires deep sample: the participants were split randomly into two groups,
examination of the model created by each subject. thus creating independent samples. The normal distribution
Each solution, within each of the two groups, was assumption was examined using KolmogorovSmirnov test
evaluated and graded by two experienced graders.2 The for each of the groups on each dependent variable (i.e.,
two graders worked and graded each solution sepa- comprehension, comprehension time, comprehension pref-
rately (using separate copies of the solution sheets). To erence, quality of model, modeling time, modeling prefer-
assist the graders, we prepared expert solutions of ence). We found out that all groups were normally
the tasks (prior to the experiment). Since there might be distributed, except for two variables: perceived compre-
other correct solutions, as discussed in Sect. 3.5, the hensibility (p value \ 0.05) and perceived quality for the
graders were instructed to use the expert solutions as OO-DFD transaction group, which has a week normal dis-
guidelines only and to grade each solution compared to tribution (p value = 0.07). Therefore, we conclude that
the requirements document. At the beginning of their t tests are suitable for all variables except these two. For
grading work, the two graders of each treatment group these, we used the Wilcoxon test, which is suitable when one
worked together on a few solutions in order to of the assumptions of the parametric test does not hold. The
synchronize their work; then they continued working homoscedasticity assumption is validated each time a t test is
separately. This way, each solution received two executed by the Levene test. When the assumption is not
independent grades. We checked the correlation fulfilled, another version of t test that does not assume the
between the two grades using Pearson correlation homoscedasticity has to be performed.
statistic, which is a mean to measure correlation among
graders when the grades are in an ordered scale.3 The
4.1 Results of the comprehension experiment

2
Because of the criticality of this grading, we decided to use two Table 7 summarizes the results of the comprehension
experienced graders for each solution. They were graduate students of experiment. It shows that there was no significant differ-
Information Systems Engineering who served as TAs of the two ence in comprehension, but it took less time to comprehend
involved courses (IS Analysis & Design, and OO Analysis & Design).
3
the UC model, while users perceived that the OO-DFD
We cannot use Kappas coefficient since it is suitable only for
model is easier to comprehend. Following the results, we
qualitative/categorical items; grades on a 0100 scale are not
categorical. accept H01 and reject H03 and H05.

123
Requirements Eng

4.2 Results of the model creation experiment no significant differences between the two methods, while
in categories C and D, there were significant differences in
As said, prior to the analysis of the model creation results, favor of the UC method. These results are reasonable
we computed the Pearson correlation between the pairs of because in OO-DFD transaction, there are more external
grades given by the two grades; we found them highly entities and data flows compared to actors and connections
correlated (p value \ 0.05), thus we could average the two in UCs; hence, there are more chances to make errors when
grades of each participant. Table 8 presents the results of using the OO-DFD transaction method, making the UC
this experiment. As can be seen, there was a significant method superior in those categories. Contrarily, in category
difference in the quality of the created models in favor of E, there was a significant difference in favor of the OO-
the UC method, but there was no significant difference in DFD transaction method, possibly because the logical
the time taken to create the models, and no difference in the descriptions of the transactions are more structured com-
users perceptions about how good are the models for pared to the informal descriptions of UCs.
modeling requirements. Following the results, we reject
H02 and accept H04 and H06.
We wanted to explore in more details the reasons for the 5 Discussion
difference in quality of the created models. For this, we
classified the possible types of errors in the diagrams and 5.1 Evaluation of results and implications
descriptions into five categories:
5.1.1 Analysis of the comprehension results
Amissing or redundant function/transaction/UC
Bmissing or wrong name of function/UC in diagram
(A) Comprehension of models. The results showed no
or description
significant difference between the comprehensibility of the
Cmissing, redundant, or wrong external entity/actor in
two methods. This result is in accord with our speculations
diagram or description
prior to the experiment due to the contrary considerations
Dmissing, redundant, or wrong direction data flow/
and the support of cognitive theories. On one hand, detailed
connection
diagrams and descriptions, as in OO-DFD transaction
Ewrong logical description of transaction/UC
versus UC diagrams and descriptions, may provide for
We counted the number of errors made by each partic- understandability, but on the other hand, too many details
ipant in each of the above categories. Then we tested the might cause confusion. As described in the hypotheses
differences in the number of errors within each of the section, the COGEVAL framework proposes cognitive
categories, using t tests. The results are summarized in principles that on the one hand support the assumption that
Table 9. As can be seen, in categories A and B, there were transaction diagrams are easier to understand than UC

Table 8 The quality of modeling results


OO-DFD UC Statistical analysis Conclusion
transactions results (a = 0.05)

Model creation grade 73.87 % 87.52 % p = 0.001 (t test) Difference in favor of UC method
Time to complete model creation task 112 min 108 min p = 0.53 (t test) No difference in time
Perceived quality of modeling method 4.04 3.92 p = 0.96 (Wilcoxon) No difference in perceived quality

Table 9 Results per categories


Category Average no of errors Average no Statistical analysis Conclusion
in OO-DFD transactions of errors in UCs results (a = 0.05)

A 1.46 1.71 p = 0.56 No significant difference


B 0.15 0.19 p = 0.81 No significant difference
C 2.38 1.05 p = 0.001 Significant difference in favor of UC method
D 3.58 0.43 p=0 Significant difference in favor of UC method
E 1.35 2.76 p = 0.014 Significant difference in favor of OO-DFD
transaction method

123
Requirements Eng

Table 10 Summary of the participants comments


Phrase No of times said No of times
about OO-DFD said about UCs
transactions

The diagram is clear, understandable and visual 8 16


The diagram contains many details and is confusing and complicated 3 0
The diagram describes clearly the data flow in the system 4 0
The transactions/use cases descriptions add information to the understanding of the system 8 3
The transactions/use cases descriptions are clear and understandable 2 10
The diagram is too general, not enough details 0 9
The descriptions are not detailed enough and it is hard to extract specific information from them 0 12

diagrams (more relationship information makes more OO-DFD transaction diagrams and descriptions include
comprehensive model), while on the other hand, they many more components and details compared to the
support the opposite assumption that transaction diagrams alternative, it takes more time to read and understand them.
are less understood (more elements in a model makes it less (C) Perceived comprehensibility of models. Surpris-
comprehensible). ingly, the participants perceive the OO-DFD transaction
In order to get more insight into the results, we analyzed model as more comprehensiblecontrary to our a priori
the comments written by the participants in the post-test assumption. This outcome is difficult to explain, in par-
questionnaires. We first reviewed the comments and ticular, if we compare it to the opposite outcome regarding
extracted phrases that appeared many times, that is, written the time dimension. A possible explanation may be that the
by many participants. Then, we counted the number of subjects thought that since the method is more detailed and
times each phrase was written about each of the two the descriptions are more structuredit must be more
methods. The results of this analysis, which are summa- comprehensible. Another possible explanation for this
rized in Table 10, support our earlier explanations. As can result might be that each participant only assessed one of
be seen, the phrase The diagram contains many details the two modeling techniques; therefore, the difference in
and is confusing and complicated was written 3 times perceived comprehensibility of the models is threatened by
about the OO-DFD transaction method, but never about the the variation in the individual appreciation of comprehen-
UC method, while the phrase The diagram is clear, sibility. What is difficult or easy for some can be
understandable and visual was written 16 times about the considered differently for others. We address this limitation
UC method and only 8 times about the OO-DFD transac- in Sect. 5.2Threats to validity.
tion method. On the other hand, the phrases The diagram
is too general; not enough details and The descriptions 5.1.2 Analysis of the model creation results
are not detailed enough and it is hard to extract specific
information from them were written many times about the (A) Quality of created model. We found a significant dif-
UC method (9 and 12, respectively), but never about the ference between the qualities of the created models in favor
OO-DFD transaction method. Table 10 also shows some of the UC method. As discussed earlier, these findings can
contradictions: the phrase The transaction/use case be explained by cognitive theories: according to the
descriptions are clear and understandable was written COGEVAL framework, when a model requires a greater
only 2 times about the OO-DFD transaction and 10 times number of simultaneous items to create a diagram, the
about the use cases, while the phrase The transaction/use diagram is of less quality than a diagram of a model which
case descriptions add information to the understanding of requires less number of items. Indeed, an OO-DFD trans-
the system was written 8 times about the OO-DFD action diagram includes more types of components com-
transaction and only 3 times about the UCs. These con- pared to a UC diagram, and an OO-DFD transaction
tradicting comments too support the result that we found, description too includes more types of components com-
that is, that there is no difference in comprehension pared to UC descriptions. Moreover, the process logic of an
between the two modeling methods. OO-DFD transaction is described in pseudo-code, while a
(B) Time to comprehend models. We found that it takes UC description is less structured, more natural-like
less time to comprehend a model expressed in a UC model. description. Hence, an OO-DFD transaction modeler has
This finding is also in accord with our assumptions prior to many more chances to commit errors compared to a UC
the experiment and can be explained as before: because modeler.

123
Requirements Eng

But note that we found significant differences in certain description of transactions, but retaining the structured
types of categories of the models: in category C (missing, description of the process logic of the transaction. The UC
redundant, or wrong external entity or actor in diagram or method can be improved by adding these missing compo-
description) and in category D (missing, redundant, or nents to the UC diagram and referring to them in the UC
wrong direction of data flow/connection), we found description.
advantage to the UC method; this is reasonable as already In the following, we briefly describe an enhanced UC
explained, because the OO-DFD transaction method method that we term EUCEnhanced Use Case, which
requires to define external entities, data classes, and data combines the advantages of the two:
flows, which are not required in the UC model. Contrarily,
A. A EUC diagram will include components that exist in
in category E (wrong logical descriptions of transaction/
an OO-DFD transaction diagram, that is, one or more
UC), we found an advantage to the OO-DFD transaction
functions, external entities signifying sources of input
method. The reason for this result is that the OO-DFD
data and destinations of output, data classes from
transaction description is more structured, thus causing
where the functions can retrieve data or where they
fewer mistakes.
can store data, and directed data flows between the
(B) Time to create models. Contrary to our a priori
respective components. Proponents of existing UC-
assumptions, we found no significant differences in time
driven methodologies may ask: Where can the data
to create the models. This result is somewhat surprising
classes come from at this stage? Two answers are
because it is expected that a more demanding task
possible: (a) if we also adopt the FOOM approach,
(transactions) would take more time. A possible expla-
prior to creating the EUCs, we create an initial class
nation for this indifference is that the participants, who
diagram and use these classes in the respective EUC
were allocated the same amount of time to work on the
diagrams; (b) even if we do not adopt the FOOM
two tasks, were advised to use the first hour for the
approach, while creating a EUC diagram, we may
comprehension task, and therefore, they have reserved
expect the modeler to define not only the functions of
about the same amount of the remaining time to work on
the EUC but also the required data classes, as well as
the model creation task.
the external entities.
(C) Perceived quality of modeling methods. We found
B. There is no need to include the traditional actors in
no significant differences in the perceived quality of the
the EUC diagram; we can list them in the description
two modeling methods. This result is somewhat inconsis-
of the EUC, where we also define other things that are
tent with the equivalent outcome regarding perceived
not part of the diagram (e.g., pre- and post-conditions).
comprehensibility of the models, where we found a sig-
Another reason for this change is that a EUC may
nificant difference. This result can be explained by the
sometimes be operated by many different types of
contradiction in the participants opinions, as shown in
operators, each having different access privileges;
Table 10. On one hand, many subjects thought that the
there is simply not enough room to include such
OO-DFD transaction method is too detailed and confusing,
details in the diagram.
while on the other hand, many others thought the UC
C. There is no need to include the special types of uses
method is too general and not detailed enough. As
and includes UCs in the EUC diagram. Instead of
described in the perceived comprehensibility of models
them, we may have just chained functions; a data flow
section, we can also explain this inconsistent result due to
from one function to another means that the first
the different individual appreciation of comprehensibility
triggers the latter and may pass some data/parameters
by the subjects, since in our experiment settings, each
to it.
subject used and assessed only one of the two modeling
D. The description of the EUC will have to distinguish
techniques. Section 5.2 further relates to this issue.
between the external entities that are source of input
or destination of output, and the operators of the use
5.1.3 An improved method for modeling functional
case at runtime. But contrary to the current too-
requirements
structured description of a transaction, we adopt a less
structured, more natural-language-like description, as
We have seen that there are differences between the two
in traditional UC descriptions.
methods, and each has some advantages and disadvantages.
Our observations led us to conclude that each of the We complete this sketch of the enhanced method by
methods can be improved by considering the advantages of showing a possible description of a EUC. We assume that
the other. The OO-DFD method can be improved by sim- the EUC is the same as the transaction diagram in Fig. 1;
plifying the redundant, overly detailed, and structured Table 11 shows its description.

123
Requirements Eng

Table 11 The EUC description


Component Description

EUC name Assign reviewers to a submitted article


Operators PC Chair, electronic mail system
Preconditions There are submitted articles in the system that have not yet been assigned to reviewers
Postconditions The chosen article is assigned to reviewers
Main success scenario 1. The system reads the details of reviewers from class Reviewer and non-assigned articles from class
Article and displays them to the PC Chair
2. The PC Chair selects and inputs to the system an article from the list and the reviewers assigned to
review it
3. The assignment details are saved in class Article review
4. The system sends the articles details and the last date to submit review reports to the reviewers, by the
electronic mail system
Extensions None

5.2 Threats to validity Of course, counting and summing the number of correct
answers is not a perfect measure, because it gives an equal
As any quasi-laboratory experiment, this one too has lim- weight to all statements. In spite of this weakness, we argue
itations and threats to validity of the results. We distinguish that the measure of understandability we used is reasonable
between four types of threats: conclusion, internal, con- and follows the best practice in other comparative
struct, and external validity. experiments of comprehension.
Conclusion validity: Conclusion validity refers to (b) Another threat to construct validity in the model
whether the conclusions reached in a study are correct. creation task is that each subject was given a requirements
For controlled experiments, conclusion validity is directly document of a certain case study (part of the IFIP Con-
related to the application of statistical tests to the data. If ference system). This does not represent a realistic situa-
the statistical tests are not applied correctly, this is a tion: in reality, analysts usually elicit requirements from
threat to the conclusion validity. In the presented exper- real users in an interactive and iterative process. We
iment, we believe that all statistical tests were applied skipped this important aspect and gave the subjects a
correctly. Yet, the reliability of measure of quality of ready requirements document. So, it can be argued that
models might introduce some threats to validity; to we mainly measured how well analysts convert predefined
address this concern, each solution was graded by two textual requirements into diagrammatic specifications. We
independent graders. cannot be sure that the same results would be obtained had
Internal validity: Internal validity refers to the extent to the subjects been working in a more realistic environment.
which the independent variable(s) were actually responsi- But on the other hand, we have no reason to assume that
ble for the effects seen to the dependent variable. A pos- this would have affected differently the results.
sible threat to the internal validity may be that unknown (c) In both comprehension and modeling tasks, a major
factors may have had an influence on the results and threat to validity relies on the fact that each participant was
therefore put limitations on the internal validity of the asked to assess only one of the modeling methods on a 17
study. This possible threat, which exists in all experiments ordinal scale. Each individual appreciation of the level of
of this kind, was minimized by the random assignment of comprehensibility and the level of the modeling methods
the subjects into the two treatment groups. compatibility to describe functional requirements is differ-
Construct validity: Construct validity is the degree to ent. What seems to be difficult for some may seem to be
which the independent and dependent variables accurately easy for others. Similarly, a method that seems to be
measure the concepts they are supposed to measure. extremely compatible to some may seem to be not at all
Threats to construct validity have been identified in both compatible to others. The ideal way would have been to
experimental tasks: choose a within-subject experiment setting, in which each
(a) In the comprehension task, a threat to validity is the subject would have experimented with the two modeling
measure of understandability. To deal with this threat, we methods. In that case, we could have asked directly questions
used a questionnaire consisting of 22 statements, thus such as: Which of the two methods do you understand
addressing many facts appearing in the presented models. better? or Which method do you think is more suitable to

123
Requirements Eng

define functional requirements? But this would have caused 6 Conclusions and future work
other problems. One is a problem of practicality: we would
need many more participants because each participant would We compared two alternative methods for modeling the
have had to perform the two tasks using one method and then functional requirements of an information system: OO-
perform two different tasks using the other method. So we DFD transaction diagrams and descriptions versus UC
either need many more (homogeneous) participants or have diagrams and descriptions. In a controlled experiment, we
each participant perform all those tasks, which would take a examined the differences between the two methods with
double amount of time (6 h at least instead of 3 h in our respect to comprehension, quality, time, and perceived
experiment)this is not so realistic in an exam setting, comprehensibility and quality of the modeling methods.
besides the problem of getting tired. Another problem with The main result of the comparative experiment is that the
such an experiment is that we have to make sure that the quality of the models created with the UC method is sig-
different tasks given to each subject are similar in compre- nificantly better than the quality of models created with the
hensibility, complexity, etc.problems that are not easy to OO-DFD transaction method. But this result is mainly
solve. Another problem with a within-subject form of because UCs do not include components that are included
experiment is the possible bias due to the order effect (since in OO-DFD transaction, and therefore, UC analysts can
each subject has to perform each task using a different avoid creating errors that cannot be avoided by OO-DFD
method); to overcome this problem, we might need to double transaction analysts. We also concluded that the OO-DFD
the number of treatment groups. (For example, subjects in transactions are overly detailed, while the UC diagrams are
one group would perform one task using OO-DFD first and too general and not detailed enough. This led us to propose
then another task using the UC method, while subjects in an improved method to define and describe the functional
another group would perform the same tasks in the opposite requirements that is a hybrid of the two.
order). Due to these practical reasons, we chose the But note that the ECU method has not yet been tested
between-subject setting, and are aware of its limitations, empirically. In the future, we plan to evaluate the enhanced
which may explain the inconsistent and surprising results method compared to the original methods. In that research,
regarding the perceived comprehensibility and quality. we plan to repeat the experiments using tasks of different
External validity: External validity is the degree to size and complexity.
which the results of the experiment can be generalized to
the population under study and other research settings. The
following possible threats have been identified: References
(a) The case studies used in the experiments are rela-
1. Arlow J, Neustadt I (2001) UML and the unified process: prac-
tively small and may not be representative in terms of size tical object oriented analysis & design. Addison Wesley, Reading
and complexity. But this limitation is true for almost all 2. Bajaj A (2004) The effect of the number of concepts on the
controlled experiments conducted in the areas of software readability of schemas: an empirical study with data models.
engineering and method evaluation. Working Paper, University of Tulsa, OK, USA
3. Bass L, Bergey J, Clements P, Merson P, Ozkaya I, Sangwan R
(b) The subjects who participated in the experiment (2006). A comparison of requirements specification methods
were students. In the comprehension task, they played the from a software architecture perspective. Technical report, Soft-
role of users who read and evaluate analysis specifications, ware Engineering Institute, Carnegie Mellon
but actually they were not real users who evaluate speci- 4. Booch G (1991) Object-oriented design with applications. Ben-
jamin/Cummings, Menlo Park, CA
fications of real systems developed for them. However, we 5. Booch G, Rumbaugh J, Jacobson I (1996) The unified modeling
have no reason to assume that the use of students as sur- language for object-oriented development. Rational Software
rogates of real users caused any bias of the comparative Corporation
results. Similarly, in the model creation task, we used the 6. Burton-Jones A, Weber R, Wand Y (2009) Guidelines for
empirical evaluations of conceptual modeling grammars. J Assoc
same students who played the role of analysts, but they Inf Syst 10:495532
were not real, experienced analysts of real requirements. 7. Coad P, Yourdon E (1991). Object oriented analysis, 2nd edn.
But again, in spite of this limitation, we have no reason to Prentice Hall, Englewood Cliffs, NJ
assume that real analysts would perform differently with 8. Cockburn A (2001) Writing effective use cases. Addison-Wesley,
Boston
the two modeling methods. Recall that our subjects were 9. Cox K, Phalp K, Shepperd M (2001) Comparing use case writing
senior software engineering students who were trained to guidelines. In: Proceedings of the 7th international workshop on
become systems analysts. Moreover, each subject learned requirements engineering: foundation for software quality,
his/her method from an expert in that method. We note pp 101112
10. DeMarco (1978) Structured Analysis and System Specification.
again that in almost all experimental research in the area, Yourdon Press, New York
the subjects were students (as it is not feasible to conduct 11. Dori D, Goodman M (1996) From object-process analysis to
such experiment in the real world). object-process design. Ann Softw Eng 2:2550

123
Requirements Eng

12. Gane C, Sarson T (1979) Structured system analysis: tools and 25. Parsons J, Cole L (2005) What do the pictures mean? Guidelines
techniques. Prentice Hall, Englewood Cliffs, NJ for experimental evaluation of representation fidelity in dia-
13. Gemino A, Wand Y (2003) Evaluating modeling techniques grammatical conceptual modeling techniques. Data Knowl Eng
based on models of learning. Commun ACM 46:7984 55:327342
14. Hadar I, Kuflik T, Perini A, Reinhartz-Berger I, Ricca F, Susi A 26. Peleg M, Dori D (2000) The model multiplicity problem:
(2010) An empirical study of requirements model understanding: experimenting with real-time specification methods. IEEE Trans
use case vs. tropos models. SAC10 March, Sierre, Switzerland, Softw Eng 6:118
pp 2226 27. Ramsin R, Paige R (2008) Process-centered review of object
15. Hahn J, Kim J (1999) Why are some diagrams easier to work oriented software development methodologies. ACM Comput
with? Effects of diagrammatic representation on the cognitive Surv 40(1):389
integration process of systems analysis and design. ACM Trans 28. Rockwell S, Bajaj A (2005) COGEVAL: applying cognitive
Comput Hum Interact 6(3):181213 theories to evaluate conceptual models. Adv Top Database Res
16. Jacobson I (1987) Object oriented development in an industrial 4:255282
environment. In: Proceeding of the object-oriented programming 29. Rumbaugh J, Blaha M, Premerlani W, Eddy F, Lorensen W
systems, languages and applications, pp 183191 (1991) Object-oriented modeling and design. Prentice-Hall,
17. Jacobson I (1992) Object oriented software engineering: a use Englewood Cliffs, NJ
case driven approach. Addison-Wesley Professional, Reading 30. Shoval P (1988) ADISSA: architectural design of information
18. Jeyayaj A, Sauter VL (2007) An empirical investigation of the systems based on structured analysis. Inf Syst 13(2):193210
effectiveness of system modeling and verification tools. Commun 31. Shoval P (2007) Functional and object-oriented analysis and
ACM 50(6):6376 design: an integrated methodology. Idea (IGI) Publishers, Hershey
19. Kabeli J, Shoval P (2005) Comprehension and quality of analysis 32. Shoval P, Kabeli J (2001) FOOM: functional- and object-oriented
specificationsa comparison of FOOM and OPM methodolo- analysis & design of information systemsan integrated meth-
gies. Inf Softw Technol 47(4):271290 odology. J Database Manag 12(1):1525
20. Kim J, Hahn J, Hahn H (2000) How do we understand system 33. Siau K (2004) Informational and computational equivalence in
with (so) many diagrams? Cognitive integration processes in comparing information modeling methods. J Database Manag
diagrammatic reasoning. Inf Syst Res 11(3):284303 15:7386
21. Larman C (2002) Applying UML and patterns: an introduction to 34. Topi H, Ramesh V (2002) Human factors research on data
object-oriented analysis and design, and the unified process, 2nd modeling: a review of prior research, an extended framework and
edn. Prentice Hall, Englewood Cliffs, NJ future research directions. J Database Manag 13(2):319
22. Mathiassen L, Munk-Madsen A, Nielsen P, Stage J (2000) Object 35. White SA (2004) Introduction to BPMN, IBM Corporation,
oriented analysis and design. Marko Publishing, Alborg business process trends
23. Millet I, Nelson R (2007) Data flow diagram vs. use cases 36. Wohlin C, Runeson P, Host M, Ohlsson MC, Regnell B, Wesslen
student perceptions. Internatl J Inf Commun Technol Educ A (2000) Experimentation in software engineering: an introduc-
3(1):7078 tion. Kluwer, Boston
24. Moody D (2006) What makes a good diagram? Improving the 37. Yu E (1997) Towards modeling and reasoning support for early-
cognitive effectiveness of diagrams in is development. In: Pro- phase requirements engineering. In: 3rd IEEE international sym-
ceedings of the 15th international conference in information posium on requirements engineering, Annapolis, USA, pp 226235
systems development (ISD)

123

Você também pode gostar