Decoding Dreams and Thoughts using AI Sentiment Analysis

Understanding and classifying the emotional content of dream descriptions is a unique challenge that sits between psychology and natural language processing. This thesis tackles this issue by creating a machine learning pipeline that can automatically predict sentiment labels from free-text dream narratives. The main method uses handcrafted lexical-semantic features for sentiment classification. Specifically, Term Frequency–Inverse Document Frequency (TF-IDF) extracts lexical patterns, while Empath, a psychologically grounded lexicon, provides high-level semantic signals. To address extreme class imbalance in the labeled data, a targeted resampling strategy based on SMOTE is used to ensure balanced representation across all sentiment categories. A Multi-Layer Perceptron (MLP) classifier is trained on the engineered features and evaluated using both 5-fold stratified cross-validation and Leave-One-Out Cross-Validation (LOOCV). The model achieves a high accuracy of 86\% and strong macro F1-scores across folds and samples. This shows robust generalization with minimal overfitting. In addition to this main pipeline, the thesis also looks into fine-tuning large language models. A DeBERTa transformer was adapted using parameter-efficient LoRA (Low-Rank Adaptation), allowing for modeling emotional language in context. Although this was not integrated into the main system, this parallel investigation highlights the potential of transformer-based models for understanding emotions in narrative text. Beyond model performance, the thesis examines how consistent the sentiment labels are by comparing multiple sources. Cross-dataset accuracy matrices and correlation analyses (Cohen’s Kappa and Spearman’s rho) show significant disagreements among annotators. This highlights the subjective nature of emotional labeling in dreams. Overall, this work provides a reproducible affective computing pipeline for narrative data and offers insights into the challenges of subjectivity in human-labeled sentiment datasets, especially in areas like dream interpretation.