EA-Miner: Towards Automation in Aspect-Oriented Requirements Engineering - Transactions on Aspect-Oriented Software Development - page 24

• Identify the AORE abstractions (such as viewpoints and early aspects); and

• Structure the requirements according to the viewpoint-based AORE model (e.g.,

creating a model such as the one in Fig. 11).
In the case of the tool the identification part considers the time to parse the input and
show the results on the screen which is a fully automated task. The structuring considers
the time to apply some filters (following the guidelines in Sect. 4.3 which are very
straightforward) and generate the final specification. The results are shown in Table 2.

Table 2. Time spent In minutes

Light Control Library

Manual EA-Miner Manual EA-Miner

Identification 49 1.03 81 1.48

Structuring 833 6 1377 10.67

The results about the total time spent show that the EA-Miner approach was
approximately 130 times faster for the light control system and 120 times faster for the
library system. The explanation for the fact that the manual approach took so long is that
the engineer doing the tasks manually had to read the whole document to identify the
concepts. Moreover, this engineer also wasted precious time to structure the final
specification document since s/he read, several times, parts of the document again and
again to copy and paste text. On the other hand the engineer doing the task with EA-
Miner had the concepts identification as well as a suggested structure fully automated,
therefore s/he had just to apply some filters and guidelines (e.g., filtering viewpoints by
their relative frequency, and grouping viewpoints synonyms and with same root form—
stemming) and then automatically generate the specification document.
One important issue to point out about EA-Miner's structuring time spent on
Table 2 is that the filtering operations offered by the tool (e.g., frequency filtering for
listing the 10 or 20 most relevant viewpoints) only takes a few seconds. Therefore the
times shown also include a bit of interaction with the user to discard and requirements
viewpoints (using the guidelines in Sect. 4.3) and generate the specification.
Another important clarification relates to the time spent to update the lexicon file
used for early aspect identification. Before running the tool with both examples (Light
control and library system) we had already an initial version of the lexicon previously
built manually while the tool was developed and tested. The time spent to build this
initial lexicon was not taken into account in our evaluation as our goal was to measure
identification and structuring activities. We do not consider this a problem as the
lexicon update process can be automated and reuse existing knowledge (e.g., NFR
catalogues) as we explained in Sect. 4.1.3.

5.1.2 Results of Accuracy
Accuracy is measured by precision and recall as below:
•

The precision for a technique Pt = (number of correct candidates identified by t ) /
(total number of candidates identified by t).

• The recall for a technique Rt = (number of correct candidates identified by t ) /

(total known correct candidates).