Abstract: To the weight update of Latent Semantic Analysis（LSA） model，this paper proposes an adaptive weight update algorithm based on Bayesian theory（ALSAB）.ALSAB adopts Maximum A Posteriori（MAP） provability estimation and Expectation-Maximization（EM） algorithm to estimate the weight parameters of LSA，and ALSAB employs incremental learning to decrease accumulative effect caused in continuous update with considering that the probability of uncommon words decreases in continuous update.Experimental results show that，compared with the existing algorithms，the proposed ALSAB algorithm greatly improves recall and precision rates of information retrieval systems.