The invention provides a conversation psychological state assessment method based on multi-
modal data. The method is applied to scenes such as century review conversation and the like. Multi-
modal data such as micro expressions, voices and intonations, physiological parameters and semantic contents of a talked object are obtained through videos, microphones and non-contact
physiological monitoring; fusing the multi-
modal features with the individual feature vectors, inputting a psychological
state recognition model, and recognizing the emotion category, the psychological elasticity level and the risk evolution trend of the talked object; further, based on cross analysis of the psychological elasticity index and semantic emotion, dynamically calculating a psychological load value, judging a psychological tolerance boundary and outputting a safety level; and finally, according to the
risk type, the change rate and the
context sensitivity, forming a three-level intervention suggestion of
risk type, evolution rate and context association. Real-time evaluation and
trend prediction of the psychological state of the conversation object are achieved, the method has the advantages of being non-contact, intelligent and high in
interpretability, and the safety and scientificity of the conversation process are effectively improved.