Ambiguity Detection in Requirements Classification Task using Fine-Tuned Transformation Technique

Abstract

The software requirement specification (SRS) document is essential in software development. This document influences all subsequent steps in software development. Nevertheless, requirements problems, such as insufficient or ambiguous specifications, can cause misunderstandings during the requirement analysis stage. This influences testing activities and increases the project’s duration and cost overrun risk. This paper represents an intuitive approach to detecting ambiguity in software requirements. The classifiershould learn ambiguous features and characteristics extracted from the text on a training set and try to detect similar characteristics from a testing set. To achieve this, this study experimented with two main approaches. The first approach is feature extraction, which uses the hidden states as features and trains asupport vector machine (SVM) classifier to assess software requirement ambiguity without modifying the pre-trained model. Unfortunately, this approach only identified 68% of the requirement ambiguity. The second approach is training an end-to-end model that updates the parameters of the pre-trained model. This approach enhanced the baseline results by 13%.