Robust Requirements Analysis in Complex Systems through Machine Learning - Trustworthy Eternal Systems via Evolving, Software Data and Knowledge - page 54

set of lexical and grammatical features expressed in the resulting weighted vector
that establishes a variety of latent semantic relations between requirements: the
closer two requirements are in the representation space, the stronger is their
semantic relation.

Fig. 3. Conceptual Graph employed to evaluate the semantic similarity function

A first use of this representation is the Graphical User Interface (GUI) to
the database, that allows to analyze complex semantic relationships between
individual requisites, such as the redundancy. As shown in Figure 3, individual
requisites are represented through a conceptual graph where edges between ver-
tices express weighted semantic similarity relationship between two instances.
In Figure 3 the graph of requisites closer to the requisite CMS-OPR-3333 ,
i.e. “ The CMS shall monitor the equipment Status through reception, extraction
and display of the information that is periodically transmitted by the surveillance
radar ... ” is shown. Notice how the most similar text is CMS-OPR-33355 :
“ The CMS shall monitor the Health Status through reception, extraction and
display of the information that is periodically transmitted by the surveillance
radar. ”. This confirms that the captured notion of similarity well reflects rich se-
mantic relations. This relationship instance is in fact a form of textual entailment
[18], i.e. the directional relationship between a text pair ,madeby T , i.e.

the entailing “ Text ”, and H , i.e. the entailed “ Hypothesis ”. It is usually stated
that T entails H if a human that reads T (assuming it to be true) would accept
that H is most likely true. This definition is somewhat informal but model an
underlying useful form of commonsense knowledge for human expert.
The way a graph is built depends on the distance metrics established within
the underlying vector space. Given a requisite r i , the short texts describing
functionalities f j can be ranked according to their semantic similarity with the
specific r i , modeled through the cosine similarity sim between the corresponding
vectors r i and f j : sim ( r i ,f j )=

T,H

r i ·

f j

r i ||·||

f j ||

||