Abstract: As the number of topics k should be given in advance in LDA model, therefore, to choose the optimal number of topics k cyclical calculation is needed, which will increases the complexity of the algorithm. To solve the problem of choosing the optimal topics k, the LDA topic increments training algorithm is proposed. First, the entropy of word-topic probability is taken as the extract standard of fuzzy word in the iterative process, and a new topic is generated by the extracted fuzzy word. Second, the dimension of the global parameter β (word-topic probability matrix), dirichlet parameter α, and number of topics k, are increased in the process of variational inference. Third, the variational training algorithm is executed with the transformed global parameters β, α and k. Finally, the LDA topic increment training algorithm is executed cyclically, which is stopped while the likelihood is converged to complete the increment training of k. The effectiveness and feasibility of the proposed algorithm for optimal topics number k is verified by experimental analysis on real datasets.

Key words: artificial intelltgence, LDA, variational inference, increments training, topic classification, natural language processing