DreamNet: A Multimodal Framework for Semantic and Emotional Analysis of Sleep Narratives

Dreams are a fundamental aspect of human experience, providing deep insights into cognitive processes, emotional states, and subconscious mechanisms. Psychological s tudies have long recognized their diagnostic value, linking dream content to conditions such as anxiety, depress ion, and post-traumatic stress disorder (PTSD), while also emphasizing their role in creativity and problem-solving [10, 12]. However, computational analysis of dream narratives has lagged behind physiological sleep re search, which employs tools like elec-troencephalography (EEG) and polysomnography to map sleep sta ges or detect anomalies [15]. Traditional methods--such as manual coding via the Hall/Van de Castle system [3, 5]--or subjective interpretation are time-consuming, bias-prone, and impractical for large-scale s tudies. Although physiological data offers precision, it overlooks the subjective narrative content that refl ects an individual's inner world. To address this gap, we present DreamNet, a multimodal deep learning framework that extracts semantic themes (e.g., flying, falling, pursuit, loss) and emotional stat es (e.g., fear, joy, anxiety, sadness) from textual dream narratives, optionally augmented with REM-st age EEG signals. Built on RoBERTa [9], a transformer model optimized for contextual understanding, DreamNet excels in text-only mode and achieves enhanced precision with physiological integration, leveraging advances in multimodal NLP [14].