A Transformer Based Approach for Classifying Software Requirements

Abstract

Optimizing and automation of the software development lifecycle (SDLC) has been a challenge in the domain of software development for a while. One of the main challenges in automation of the SDLC lies within the work of collecting, analyzing and establishing requirements. Requirements that are collected from the clients are unstructured and often noisy, which takes substantial amount of effort to sort out and clarify to create a well organized Specification Document (SRS Document). To address this challenge, we propose a transformer based ensemble approach to classify the requirements so that the work of establishing requirements from client descriptions can be significantly accelerated. Our approach involves adopting the DistilBERT and RoBERTa models, which are known for their lightweight performance and combine them with ensemble models for classification. One of our considerations was the cost of computing resources, which is why we chose lightweight transformer models that are not as resource intensive like some other state of the art language models. Our experiments compares our approach with out of the box classifiers that are available with DistilBERT and RoBERTa as well as state of the art GPT model. The current experiment shows that our proposed approach outperformed out of the box classification capabilities offered by DistilBERT and RoBERTa in terms of accuracy, which is a widely adopted metric in terms of classification tasks. Our work aims to contribute to the development of lightweight classifiers that would aid in the task of classifying software requirements. Furthermore, our study provides insights to how the state of the art GPT models perform in this task of classifying software requirements. We hope our work will contribute to the future advent of automation of the SDLC.