Abstract: Non-functional requirements (NFR) describes a set of quality attributes required by the software, such as safety, reliability, performance, etc. In order to develop high-quality software products, it would be beneficial to automatically extract NFR from the Software Requirements Specification (SRS), which not only reduces the labor, time, and mental fatigue involved in identifying specific requirements from a large number of requirements, but also helps developers provide high-quality software that fully meets user expectations. In order to solve this problem, by adopting deep learning feature extraction and classification technology, a BERT-LSTM network model based on the combination of pre-trained BERT word embeddings and long short-term memory network LSTM is proposed, which is used for automatic NFR classification of quality software development. First, use the BERT model to train the word vectors in the sentence. Then use the Long Short-Term Memory network to further perform feature selection and dimensionality reduction. And finally use the Softmax classifier to identify NFR from SRS. Experiments show that in the PROMISE corpus composed of NFR, the BERT-LSTM network model has achieved the best results compared with other algorithms in terms of precision, recall, F₁ score, and other indicators.

Key words: non-functional requirements, software requirements specification, BERT, long short-term memory