At present, most of the researches of sentiment classification are carried out from the writer's perspective with quite few analyses from readers.This paper is to study the sentiment analysis from the news readers.A model of multi-label correlation sentiment classification based on completion matrix and LDA is proposed to extract the topic.The original news text is represented with the generated text-subject features, which are taken as the input to a subsequent classifier.Furthermore, the paper constructs a model of enhanced completion label matrix (CM-LDA) by appending the label correlation matrix to the original label matrix.Results show that the accuracy of this approach achieves 85.72% in the multi-label classification task, which outperforms the traditional LDA methods significantly.