Deep learning-based multimodal emotion recognition from audio, visual, and text modalities: A systematic review of recent advancements and future prospects

Emotion recognition has recently attracted extensive interest due to its significant applications to human–computer interaction. The expression of human emotion depends on various verbal and non-verbal languages like audio, visual, text, etc. Emotion recognition is thus well suited as a multimodal rather than single-modal learning problem. Owing to the powerful feature learning capability, extensive deep learning methods have been recently leveraged to capture high-level emotional feature representations for multimodal emotion recognition (MER). Therefore, this paper makes the first effort in comprehensively summarize recent advances in deep learning-based multimodal emotion recognition (DL-MER) involved in audio, visual, and text modalities. We focus on: (1) MER milestones are given to summarize the development tendency of MER, and conventional multimodal emotional datasets are provided; (2) The core principles of typical deep learning models and its recent advancements are overviewed; (3) A systematic survey and taxonomy is provided to cover the state-of-the-art methods related to two key steps in a MER system, including feature extraction and multimodal information fusion; (4) The research challenges and open issues in this field are discussed, and promising future directions are given.