Enhancing Software Requirements Classification Using AI-Based Text Processing Techniques

Abstract

The categorization of software requirements as functional (FR) and non-functional requirements (NFR) is an important problem in software engineering that is widely performed using manual analysis by domain experts. The process is tedious, prone to errors, and lacks consistency from one project to another and from one organization to the next. New approaches are needed to address this challenge, we present such an approach, a new hybrid approach, which combines pre-trained transformer models to automate the requirements classification process, and machine learning techniques. We propose a novel methodology that utilizes BERT (Bidirectional Encoder Representations from Transformers) for contextual feature extraction, supplemented with multi-head attention mechanisms and bidirectional LSTM layers for capturing sequential dependencies in requirements text. We combine this deep learning architecture using ensemble classifiers (Support Vector Machines and Random Forest) with a weighted voting mechanism. We perform an experimental validation using two of the most common datasets (PROMISE with 625 requirements and PURE with 969 requirements), showing that our approach can achieve 94.2% accuracy and 94.1% F1 score for binary classification and an average of 89.7 F1 score across six categories (Security, Performance, Usability, Reliability, Portability, and Maintainability) for the multi-class NFR categorization task. Statistical significance tests further confirm that our hybrid model substantially outperforms each of the state-the-art approaches, with absolute gain ranging from 2.1% to 14.4% respective to different evaluation criteria. This directly addresses some of the main challenges in automated requirements engineering, and hence a unique, pragmatic approach towards large-scale software development processes.