A newly proposed explainable parallel multi-transformer ensemble framework significantly enhances the accuracy of effort estimation for user stories in agile software development. To address this widespread issue, researchers developed the Parallel Multi-Transformer Ensemble (PMTE) framework to automate and refine story point predictions directly from unstructured software requirements text.
A primary contribution of this research work involves the curation of a dedicated benchmark dataset comprising 4079 unique user stories gathered from various open source software projects. Because raw project text gathered from open repositories often suffers from data noise and data leakage, industrial Scrum experts designed and executed a structured three-stage curation protocol.
Following the established Design Science Research methodology, the authors engineered an integrated hybrid parallel framework capable of extracting complementary linguistic features. The proposed architecture combines two pre-trained transformer encoders, Bidirectional Encoder Representations from Transformers (BERT) and Distilled Robustly Optimized BERT Approach (DistilRoBERTa), with an optimized Term Frequency-Inverse Document Frequency (TF-IDF) keyword frequency branch. This parallel multi-path design simultaneously captures deep bidirectional contextual semantics and critical local keyword frequencies from raw story text.
These fused semantic-frequency feature representations are subsequently fed into a bidirectional Long Short-Term Memory (BiLSTM) network and projected onto a regularized regression head.
Standard regression metrics demonstrated high predictive performance on an independent test set. The framework achieved a continuous Mean Absolute Error (MAE) of 0.5124, an agile Fibonacci-mapped MAE of 0.4443, and an R2 of 0.9283. These experimental outcomes confirm the model’s superior capability in learning directly from unstructured text, providing precise quantitative estimates for agile teams.
To assess external generalizability, the researchers evaluated the trained framework on an external, independent dataset consisting of 514 user stories collected from a private software house. The model maintained stable performance across this proprietary industry sample, confirming that the parallel feature extraction pipeline generalizes robustly beyond open source repositories to distinct commercial software development environments.
To resolve the black-box opacity typical of deep neural networks, the architecture integrates post-hoc explainable artificial intelligence (XAI) at the word level. This analytical layer provides software project managers with a clear and visual understanding of the model’s predictions.
The research team translated these analytical capabilities and predictive insights into a practical web-based decision support tool tailored for Scrum sprint planning and backlog grooming. By supplying an objective, reliable, and transparent data-driven baseline, the tool assists Scrum teams in organizing upcoming sprint activities. The findings confirm that combining parallel transformer representations with localized keyword frequencies substantially improves estimation accuracy in agile engineering.
Subject of Research: Agile software effort estimation using an explainable parallel multi-transformer ensemble
Article Title: An explainable parallel multi-transformer ensemble framework for agile software effort estimation
Article References: Malik, S., Hamid, M., & Saleem, M. (2026). An explainable parallel multi-transformer ensemble framework for agile software effort estimation. Journal of Big Data. https://doi.org/10.1186/s40537-026-01602-6
Image Credits: AI Generated
DOI: 10.1186/s40537-026-01602-6
Keywords: agile software development, effort estimation, user stories, PMTE framework, multi-transformer ensemble, BERT, DistilRoBERTa, TF-IDF, BiLSTM, regularized regression, explainable AI, word-level XAI, Scrum sprint planning, backlog grooming, design science research, data curation protocol, mean absolute error
News Source: Denise Maddox. (November 21, 2026). Parallel Multi-Transformer Ensemble Accurately Estimates Agile Software Development Effort. Scienmag.



