Disclosed is a method for detecting an abnormal message, comprising: diving a text of a detected message into a plurality of text segments; obtaining one or more account attributes of each text segment, and determining a publication proportion parameter corresponding to the account attributes of each text segment; determining a first factor corresponding to the account attributes of each text segment according to the publication proportion parameter; determining a second factor of the detected message according to the first factor corresponding to the account attributes of each text segment; and determining according to the second factor of the detected message whether the detected message is an abnormal message. Through the combination of publication account attributes of messages with undifferentiated text segmentation and the use of Bayesian algorithm, batches of junk messages of a microblog account are effectively limited, and the flexibility of junk message processing is improved.