What question did this study set out to answer?

April 26, 2026Open Access

Accurate and robust ambiguity detection in software requirements documents using a GloVe-BiLSTM deep learning framework with data augmentation

Key Points

The research aims to automate ambiguity detection in software requirements documents using a deep learning framework.
Developed a GloVe-BiLSTM model for ambiguity detection in natural language documents.
Incorporated data augmentation through back translation to enhance model generalizability.
Conducted an ablation study comparing GloVe_BiLSTM with other configurations like GloVe_CNN and GloVe_LSTM.
GloVe_BiLSTM achieved 92.06% accuracy and outperformed other models.
Precision was recorded at 92.04%, recall at 91.52%, and F1-score at 91.65%.
Results indicate improved efficiency and reliability in ambiguity detection for software requirements.

Abstract

Requirements engineering is an important part of software production, maintenance, and development. Requirements engineering documents are typically written in natural language, which is prone to uncertainties, and identification of these ambiguities is essential for successful software development. Nevertheless, categorizing and identifying these uncertainties by human experts is both time-consuming and expensive. Automating this procedure improves precision and lowers costs. This research proposes a novel deep learning-based ambiguity detection framework, addressing a critical challenge in software development. The proposed framework relies on GloVe word embeddings and a Bidirectional Long Short-Term Memory (Bi-LSTM) model. The reason for choosing the Bi-LSTM model is its ability to model sequential dependencies and contextual relationships in natural language. GloVe embeddings were selected because of their ability to effectively capture semantic similarities in text, which is essential for comprehending ambiguous language in requirements specifications. Our approach also incorporates data augmentation using back translation to mitigate the class imbalance by generating diverse samples of the minority class and improving the model’s generalizability. In addition to the proposed GloVeBiLSTM model, we conducted an ablation study to evaluate the impact of different components by comparing it with alternative configurations, including GloVeCNN, GloVeCNNRNNLSTM, and GloVeLSTM. The experimental dataset was a publicly available Fault-Prone SRS dataset, which contains 7061 data points across six ambiguity classes, namely, Lexical, Syntactic, Semantic, Syntax, Pragmatic, and Clean. This study provided insights into the contributions of each component, with GloVeBiLSTM outperforming other models, achieving over 90% accuracy and superior performance across precision, recall, and F1-score metrics. Specifically, GloVeBiLSTM attained a precision of 92. 04%, recall of 91. 52%, F1-score of 91. 65%, and an accuracy of 92. 06%. These results demonstrate the robustness and efficacy of the proposed approach, highlighting its potential for enhancing ambiguity detection in software requirements. The proposed approach emphasizes interpretability by identifying specific ambiguous patterns in requirements, making it practical for real-world applications. This method not only improves efficiency in requirements engineering but also offers a scalable solution with significant potential for reducing costs and increasing the reliability of requirements specification in software projects.

KI fragen

Bookmark

View Full Paper