Chinese Aligned Multimodal Sentiment Dataset v1.0

It contains natural conversational and interview scenarios, covering three modalities: text, audio, and vision. All original video and audio data have been de-identified; only sentence-level aligned feature vectors and corresponding sentiment annotations are provided.The dataset consists of 2,281 Chinese utterances, and each sample includes:Text modality: [CLS] vector extracted by BERT-base-chinese (768 dimensions);Audio modality: Prosodic and spectral features extracted using COVAREP, aggregated by temporal mean pooling into a 74-dimensional vector;Vision modality: Action Units (AU), eye gaze and head pose features extracted with Facet, mean-pooled into a 35-dimensional vector;Sentiment intensity label: ranged in [-3, +3], where negative values denote negative emotions, positive values denote positive emotions, and larger absolute values indicate stronger intensity.The dataset also provides train/validation/test splits to ensure reproducibility of experiments.

It is suitable for research on multimodal sentiment modeling, confidence-aware fusion methods, and training and evaluation of related machine learning and deep learning models.