Abstract: The present application discloses a method of training a sentiment preference recognition model for comment information, a recognition method and device thereof, which belong to the field of natural language processing technology. The method of training a sentiment preference recognition model for comment information includes training and optimizing, based on each piece of comment information, and corresponding original text template and mirrored text template thereof, a sentiment preference recognition model in the original channel and the mirror channel, so that the sentiment preference recognition model is used to output a sentiment preference result for the two comparison objects in the comment information in terms of the attribute.