The process to vectorize the requirements in order to be used for classification is implemented as follows:
- Unnecessary characters removal
- Lower-casing
- Punctuation signs removal
- Lemmatization
- Stop word removal
- Label ecoding
- TF-IDF Vectorization
Once the sentences were processed, a further analysis using the CHI squared algorithm for feature selection identified the following most correlated unigrams and bigramas per Quality Attribute:
| Quality Attribute | Unigrams | Bigrams |
|---|---|---|
| AVAILABILITY | failure, achieve, hours, availability, available |
system must, shall available |
| FAULT TOLERANCE | eg, control, result, failure, operate |
within system, system shall |
| MAINTAINABILITY | maintain, design, new, update, maintenance |
use system, user able |
| PERFORMANCE | status, result, less, response, fast |
less fast, response time |
| SCALABILITY | manner, capable, support, handle, number |
shall support, shall capable |
| SECURITY | authorize, password, security, encrypt, access |
user system, authorize user |
| USABILITY | use, content, navigation, easy, page |
shall easy, use system |
Finally, the features were split and exported with a 17:3 ratio for training and test sets, the pickles X_train, X_test, y_train and y_test are expected to be implemented on the Automated Model Configuration repository. The same split ratio and random state are going to be used for the deep learning classifiers as well.