Information Technology Reference
In-Depth Information
type. Then, the approach maps nouns to classes (in the sense of object oriented design),
and verbs to operations. The approach can handle four predefined verb classes.
Overmyer et al. developed a tool allowing the user to mark words or word sequences
and map them to classes, roles, and operations. As opposed to the approach by Saeki et
al., they do not assume that the verb must fall into one of the four predefined categories.
Ermagan et al. developed the tool SODA, allowing to link textual use cases to behav-
ior models. However, SODA sticks to manual modeling and does not provide automatic
extraction of model elements.
Our approach has the important advantage that it does not assume the textual doc-
ument to contain sufficient information to generate models. Thus, it can cope with in-
complete documents and make the incompleteness apparent.
5.2 Natural Language Processing in Requirements Engineering
There are three areas where natural language processing is applied to requirements engi-
neering: assessment of document quality, identification and classification of application
specific concepts, and analysis of system behavior.
Approaches to the assessment of document quality were introduced, for example, by
Rupp [14], Fabbrini et al. [15], Kamsties et al. [16], and Chantree et al. [17]. These
approaches define writing guidelines and measure document quality by the degree to
which the document satisfies the guidelines. These approaches have a different focus
from our work: their aim is to detect poor phrasing and to improve it, they do not target
system modeling, and do not generate any model-related feedback, as our approach does.
Another class of approaches, as, for example, those by Goldin and Berry [18], Ab-
bott [19], or Sawyer et al. [20] analyzes the requirements documents, extracts applica-
tion-specific concepts, and provides an initial static model of the application domain.
However, these approaches do not generate feedback addressing specification incom-
pleteness either.
The approaches analyzing system behavior translate requirements documents to ex-
ecutable models by analyzing linguistic patterns. Vadera and Meziane [21] propose a
procedure to translate certain linguistic patterns into first order logic and then to the
specification language VDM, but they do not provide automation for this procedure.
Gervasi and Zowghi [22] go further and introduce a restricted language, a subset of
English. They automatically translate textual requirements written in this restricted lan-
guage to first order logic. Similarly, Breaux et al. [23] introduce a restricted language
and translate this language to description logic. Avrunin et al. [24] translate natural lan-
guage to temporal logic. Our work goes further than the above approaches, as we not
only translate textual descriptions to models, but also use the resulting models to make
the deficiencies of the textual specification apparent.
To summarize, to the best of our knowledge, there is no approach to documents
analysis, yet, able not only to translate model descriptions to models themselves, but
also to generate feedback concerning the deficiencies of the document.
6
Summary
Even though many formal and semi-formal specification techniques exist, requirements
specifications in natural language remain a de-facto standard. Apart from being readable
