NLP-driven analysis of linguistic markers, readability, and textual features in communication retractions using COPE guidelines

This study investigates linguistic and textual patterns in retracted communication studies to identify potential markers of research integrity breaches. Analyzing 367 retracted articles from the Retraction Watch database, classified using Committee on Publication Ethics (COPE) guidelines, it employs computational linguistics to examine the relationship between discursive markers and COPE-classified misconduct. Specifically, it investigates whether retracted abstracts exhibit lower readability or lexical diversity, indicating epistemic instability. Employing SSCI-SciBERT embeddings, dimensionality reduction, and density-based clustering, the study assesses semantic structure across retraction categories. Readability indices, lexical diversity, and syntactic features were compared against a matched corpus of 367 highly cited, nonretracted Web of Science articles. Results indicate that retracted abstracts exhibit reduced type-token and hapax legomena ratios, increased named entity density, and higher sentence counts. Multivariate analyses reveal statistically significant differentiation across COPE-aligned categories, particularly in subjectivity and booster usage. Supervised classification models, notably Support Vector Machine and Logistic Regression (cross-validated AUC = 0.90), achieved robust discrimination between retracted and nonretracted texts within the sampled corpus, demonstrating the predictive power of abstract-level linguistic features alone. Interpreted through Hyland’s stance framework and Information Manipulation Theory, findings suggest that rhetorical imbalance and textual opacity are statistically associated with retracted communication abstracts, warranting further theoretical investigation.