This project focuses on modeling emotional states in multimodal conversations by integrating information from text, audio, and visual modalities. The core hypothesis is that different modalities may convey the same, similar, or even conflicting emotional states, and that their contributions to the overall multimodal emotional understanding are unequal and context-dependent. To explore this, the project will: * Investigate how emotions are expressed and perceived differently across modalities in real-world conversational data. * Develop models that can identify and align modality-specific emotional cues, including cases of redundancy, complementarity, and contradiction. * Propose novel fusion strategies that dynamically weigh the contributions of each modality based on context, speaker identity, and conversational dynamics. * Evaluate the models on benchmark multimodal emotion datasets and in downstream applications such as empathy-aware dialogue systems and mental health monitoring. This research aims to advance the understanding of multimodal emotion dynamics and to build emotionally intelligent systems that are more robust, explainable, and socially aware.