A Hybrid Deep Learning Framework for Scientific Correction Text Classification

Abstract

Ensuring research integrity and maintaining the reliability of scientific communication requires a clear and comprehensive understanding of the underlying causes of article correction reasons across domains. Correction notices issued to authors are often unstructured with inconsistent textual information, making their automatic classification a challenging task. Existing approaches primarily rely on manual analysis or conventional machine learning techniques that fail to capture the contextual semantics and complex linguistic patterns present in the given correction statements. To address this limitation, this study proposes a deep learning-based framework for the automated classification of correction reasons using a large-scale dataset. The approach integrates contextual semantic representations with a hybrid classification strategy that combines data normalization, ensemble learning, and soft voting mechanisms to enhance classification performance and overall predictive stability. The experimental results demonstrate that the framework outperforms conventional machine learning and existing deep learning algorithms across multiple evaluation metrics. The model achieves an accuracy of 94.3%, along with strong specificity and predictive reliability. The model consistently demonstrates strong evaluation metrics, such as precision, recall, and F1-score, showcasing its effectiveness in handling a complex and imbalanced retraction dataset. The experimental analysis was conducted on a dataset categorized into four distinct classes, namely Correction, Expression of Concern, Reinstatement, and Retraction. The findings of the study offer a scalable and robust solution for systematically identifying correction reasons and uncovering underlying patterns. By enabling transparent, consistent, and efficient interpretation of correction information, this proposed approach can support publishers, academic institutions, policymakers, and researchers in detecting the patterns of scientific misconduct and thus facilitating informed decision-making while mitigating the risk of misinformation in scientific communication.