Auto-redirect paused manually.
ARTIFICIAL INTELLIGENCE STRATEGIES FOR AUTOMATED SOFTWARE REQUIREMENT CLASSIFICATION TO SUPPORT ARCHITECTURAL DESIGN
Abstract
Abstract
Software Requirements Engineering Is Important To Software Development Because, According To The Literature, 70-80% Of System Failures Are Because Software Requirements Are Poorly Articulated, Non-Existent, Or Incomplete. Manually Classifying Requirements Into Functional (Fr) And Non-Functional (Nfr) Categories Is Labor Intensive, And In Large-Scale Projects, It Takes About 40–60% Of The Total Requirements Engineering Resources. This Research Presents An Automated Requirements Classification Model That Adopts A Transformer-Based Model Using Bert (Bidirectional Encoder Representations From Transformers). This System Is Trained With The Limsc/Fr-Nfr-Classification Dataset, Which Contains Over 10,000 Software Requirements With Classifications, And Is Publicly Available. With This System, Software Requirements Can Be Classified With An Accuracy Of 93%, Precision Of 91%, Recall Of 89%, And F1-Score Of 0.90. Tokenization Is Implemented With A Maximum Length Of 128 Tokens, And The System Is Trained For 3 Epochs With A 16 Size Batch. This Approach Can Reduce About 65-75% Of The Manual Classification Requirement, And Thus Aid Scalability Of Software Engineering In The Industry. Additionally, This Research Goes Beyond Classification And Allows Initial Mapping Of Requirements To Design Logic, Which Supports Better Design Decisions Of A System
Software Requirements Engineering Is Important To Software Development Because, According To The Literature, 70-80% Of System Failures Are Because Software Requirements Are Poorly Articulated, Non-Existent, Or Incomplete. Manually Classifying Requirements Into Functional (Fr) And Non-Functional (Nfr) Categories Is Labor Intensive, And In Large-Scale Projects, It Takes About 40–60% Of The Total Requirements Engineering Resources. This Research Presents An Automated Requirements Classification Model That Adopts A Transformer-Based Model Using Bert (Bidirectional Encoder Representations From Transformers). This System Is Trained With The Limsc/Fr-Nfr-Classification Dataset, Which Contains Over 10,000 Software Requirements With Classifications, And Is Publicly Available. With This System, Software Requirements Can Be Classified With An Accuracy Of 93%, Precision Of 91%, Recall Of 89%, And F1-Score Of 0.90. Tokenization Is Implemented With A Maximum Length Of 128 Tokens, And The System Is Trained For 3 Epochs With A 16 Size Batch. This Approach Can Reduce About 65-75% Of The Manual Classification Requirement, And Thus Aid Scalability Of Software Engineering In The Industry. Additionally, This Research Goes Beyond Classification And Allows Initial Mapping Of Requirements To Design Logic, Which Supports Better Design Decisions Of A System