The invention discloses a semantic analysis method and belongs to the field of a computer. The method comprises the following steps of: acquiring a training corpus; segmenting the training corpus into a*M*M data blocks according to document dimensionality and word dimensionality; performing iteration sampling to the a*M*M segmented data blocks and outputting the finally converged sampling model; according to the finally converged sampling model, performing semantic analysis. Through the semantic analysis method, the acquired training corpus is segmented into a*M*M data blocks all which are executed in iteration sampling to output the finally converged sampling model and analyzed according to the finally converged sampling model. Because the segmented data blocks can be independently sampled, the memory requirement of the single process is reduced, and the semantic analysis method can solve the problem that the memory requirement of the single process memory is higher in the existing technology when there are more to-be-trained latent semantics, thereby improving the training efficiency.