Custom Data Augmentation Technique (a Deeper Insight)

This paper presents a deeper insight into prior research for the classification of NFRs based on deep learning techniques. This classification system incorporates a novel custom data augmentation approach that preserves domain vocabulary. Previously, for DNNs training, the entire corpus was enhanced before being separated into train/validation sets. This paper adopts a contrasting approach to train four deep neural networks with only augmentation on the train set. We compare our method to a baseline (no augmentation) and a state-of-the-art EDA (Easy data augmentation technique) that uses pre-trained word embeddings. Our findings indicate a marginal difference when trained with our proposed approach compared to the EDA under these settings. Additionally, with augmentation on both the train and validation set, the CNN results improved by 10% Overall.