Full Program »
A Deep Learning-Based Framework For The Classifi-Cation of Non-Functional Requirements
State-of-the-art solutions to the classification of Non-functional re-quirements (NFRs) are mostly based on supervised machine learning models, which require to invest a reasonable amount of time in feature engineering. Deep learning, on the other hand, does not need to define features explicitly. The objective of this research is to design and de-velop an automatic system to classify Non-functional requirements in multiple classes based on deep learning techniques. We used skip-gram as feature embedding and four neural networks; ANN (Artificial Neural Network), CNN (Convolutional Neural Network), LSTM (Long Short-term Memory) and GRU (Gated Recurrent Unit) to design the framework. However, these models require large, annotated corpus and overfitting is another challenging issue. To address this, we pro-posed a novel framework for text augmentation called CUSTOM data augmentation. This technique uses the sort and concatenates ap-proach to merge two sentences belonging to the same class to produce more samples yet preserving the domain vocabulary. Further to avoid overfitting we used this approach in combination to a skip-gram model pre-trained on Eng-CoNLL corpus and later fine-tuned on our corpus. We have compared our results with state-of the art (Easy data augmentation) EDA approach. Our findings indicate that deep learning model when trained with fine-tuned word embedding and CUSTOM augmentation approach improved upon the results from earlier experiments. Interestingly CNN turned out to be an outstanding learner with the jump in accura-cy from 60% to 96% when compares with the first approach.