Robust Requirements Analysis in Complex Systems through Machine Learning - Trustworthy Eternal Systems via Evolving, Software Data and Knowledge - page 53

of Templates , defined by the ontology, includes 65 templates that correspond to
the range of the function mapping each requisite to its corresponding template.
The high number of target class makes this task very challenging with respect
to the previous requisite identification problem.
SVM classifiers have been employed even in this task: parameter estimation
has not been employed, to prove the low dependence of the learning algorithm
from external parameters; instead, results are reported as mean accuracy and
(negligible) standard deviation in parenthesis. Requisites are here represented
similarly to the previous task, thus employing the BoW model, that consider
only lexical information, and the BoW + N-Words model that consider also the
shallow syntactic information of the requisites. In term of percentage of requisite
correctly covered by a template we have a classification results 87,61% (1,16%)
with Bow models and 88,5% (1,46%) with BoW + N-Words model. Even
in this evaluation, results show an accuracy of 88% proving the IE system as a
largely applicable process.

4.2 Retrieval in Large Repositories of Software Documentation

In this section the contribution of the proposed approach for Requirements Anal-
ysis (RA) is investigated in an Information Retrieval scenario to improve the
software reusability. In order to retrieve a piece of software or any other existing
functionality that satisfies a specific user requirement, a Requirement Analyst
usually retrieves existing documentation through a search engine through spe-
cific term-based queries. In a Ad-hoc Retrieval scenario [12], the quality of the
retrieved material is strictly dependent from the expressed query that reflects
user needs. In this section we instead define a robust search engine to enable
the Requirement Analyst the retrieval of existing software functionalities by ex-
pressing software requirements in natural language.
The contribution of the proposed architecture is shown here to enable a more
conceptual kind of search. The idea is that requirements determine complex
queries can be processed by our RA software and used to retrieve existing soft-
ware compatible with the functionalities expressed by the user requirements.
More formally, the user express a requisite r i ∈ R in order to retrieve one
of the specific functionalities f 1 ,...,f n i satisfying r i .Wedenotethesetofall
{f j |j =1 , ..., n i } .Asanexample,givenarequisite r “ The CS

shall provide facilities for Human Computer Interface presentation ”, we would
like to retrieve the implemented functionalities satisfying the specific need, such
as the functionality f “ The CMS shall provide the following facilities at CMS
consoles : screens, pointing device, keyboards, MFKA, service settings. ”. In fact,
in this case, the requisite r is satisfied by f , because it expresses the avail-
able facilities to interact with the software system. We collected all the pairs
RF =

as F i and n i = |F i |

,where r i is a system requirement and F i is the set of the
corresponding functionalities satisfying r i , denoted by

{r i ,F i }

{f 1 ,f 2 , ... , f n i }

.
To associate a generic f toagiven r , we first exploit the vector representation
described in Section 4.1 that reflects the generic notion of semantic text similarity
[16,17]. This is a geometric representation of textual meaning based on the the