This thesis presents a multidimensional analysis of the relationship between sentiment as evaluated by a large language model (RoBERTa) and self-reported emotional states during mind-wandering episodes. The study utilized data from the Audiovisual Attention (AVA) study, which captured the spontaneous speech of participants. The RoBERTa model was employed to quantify the emotional valence of this speech, categorizing it into positive, negative, or neutral sentiments. In parallel, self-reported measures were taken, allowing for a comparison between computational sentiment analysis and participant self-assessment. Findings demonstrate that while both approaches provide valuable insights, they offer complementary information regarding the emotional valence of thoughts. The research highlights the prevalence of neutral sentiments in mind-wandering and the model’s accuracy in classifying emotional tones, suggesting the potential of natural language processing tools in psychological research.