Detection of Cognitive Distortions in Students’ Thoughts Using Topic Modeling and Fuzzy Clustering

Abstract

Cognitive distortion (CD) refers to an irrational thinking pattern that causes individuals to misinterpret information and convince themselves of incorrect information. This study investigates patterns of CDs in students’ thoughts after academic exams. Machine learning models are utilized to detect and categorize these distortions. The methodology of this study utilizes topic modeling with two approaches: latent Dirichlet allocation (LDA) with Tf–Idf and non-negative matrix factorization (NMF). The NMF approach is applied with two different pre-trained embeddings (AraBERT and AraGPT). Fuzzy clustering is combined with these topic modeling approaches, and the results are compared. Experiments are conducted using two datasets: a collection of students’ thoughts and a generated dataset that is based on cognitive behavioral therapy (CBT) principles. When analyzing the students’ thoughts dataset, NMF with AraBERT demonstrated superior performance by producing the most meaningful topics with a coherence score of 0.78. However, in the generated dataset, NMF with AraGPT achieved a better balance between coherence and separation, along with clearer topic boundaries. Although NMF with AraBERT achieves the highest coherence score (0.86), it shows significant topic overlap inferred from the inter-clustering score (0.81). Fuzzy clustering, topic modeling, and NMF-AraGPT together provide the highest overall performance when applied to the students’ dataset. This combination provides distinct and well-separated topics inferred from the inter-clustering score (0.53). NMF topic modeling with AraGPT is the most effective model when integrated with fuzzy clustering based on the comprehensive analysis.