TITLE:
Ultra-Fast Cognitive Distortion Classification from Short Text-A Lightweight TF-IDF and Logistic Regression Pipeline on Synthetic Data
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
Cognitive Distortions, Mental Health, Text Classification, TF-IDF, Logistic Regression, Synthetic Data
JOURNAL NAME:
Open Access Library Journal,
Vol.13 No.3,
March
13,
2026
ABSTRACT: Automatic detection of cognitive distortions from short written text could support large-scale mental-health screening and digital cognitive-behavioural therapy (CBT). Many recent approaches rely on heavy deep-learning models and large datasets that are difficult to deploy in real time or in resource-constrained settings. This work presents a compact, fully transparent pipeline for binary and multi-class classification of cognitive distortions using a small synthetic corpus of brief statements. We simulate 300 short texts covering 10 canonical distortion types (e.g., all-or-nothing thinking, overgeneralization, labelling) plus neutral statements. Texts are vectorized with a 100-dimensional TF-IDF representation over uni and bi grams, and three lightweight classifiers are compared: logistic regression, random forest, and linear SVM. On a stratified 60/20/20 train validation test split, logistic regression and linear SVM both achieve perfect test performance for the binary task (Accuracy = Precision = Recall = F1 = 1.00; AUC = 1.00), while random forest reaches 0.98 accuracy and 0.98 F1. A separate multinomial logistic-regression model trained only on distorted texts correctly identifies the specific distortion type with 0.96 accuracy across 10 classes. Five-fold cross-validation confirms the stability of the pipeline (mean accuracy 1.00, SD 0.00) on this synthetic dataset. Although the unrealistically high scores are driven by the small, highly patterned synthetic corpus, the results demonstrate that compact TF-IDF models can deliver ultra-fast, interpretable cognitive-distortion classification and provide a practical blueprint for future work on larger, clinically realistic datasets.