Aiming at the deficiency of traditional text representation model, which usually ignores term correlation, and topic drifting problem during topic tracking, this paper propose an approach called self-adaptive microblog hot topic tracking method using terms correlation. Mutual information between terms in the same and different microblogs are investigated. Then the conventional text representation model is updated. Similarity calculation is performed to decide whether it is the subsequent discussions of a certain hot topic. Finally, the vectors of microblogs are updated to avoid topic drifting. Experiments show the effectiveness of our method.