The invention discloses a training sample processing method and a training sample processing device, and relates to the technical field of computer application. The problem that the existing training SVM model is of low efficiency is solved. The method comprises the following steps: acquiring an original database, wherein the original database includes all words involved in training samples, and the training samples involve at least two different types of training samples; calculating the assumed probability of each word in the original database based on a Bayesian algorithm, wherein the assumed probability of each word is the possibility that a training sample is a type of training sample when the word appears in the training sample; extracting the words of which the assumed probability is within a preset scope to get a target database; and converting the training samples into a training sample matrix based on the words in the target database to get de-noised training samples. The training sample processing method and the training sample processing device are used in the process of training sample de-noising.