With the continuous development and improvement of textual topic modeling, variational inference, as an effective approximate inference method, is widely used in parameter estimation of topic models. In this paper, combining Bayesian network and hierarchical Delikerian process (HDP) model, an HDP online variational Bayesian inference (Dist-LDA-VB) method is proposed and applied to the task of multidimensional inference of discourse hierarchical features in English corpus. By comparing the corresponding topics derived from the two models, it can be found that they have similar thematic content. In the topic inference of the constructed English corpus, the DistLDA-VB model and Markov Topic Models (MTMs) yielded similar topics, which are suitable for corpus discourse hierarchical feature inference. In addition, based on the corpus approach, this paper explores the semantic differences between the English “conclusively infer” class of synonymous adverbs certainly, definitely, necessarily and surely. The results of the study show the spatial distance between the target words and the variables, which is helpful for learners to memorize the correspondence between them, so as to systematize the inter-word differences and reduce the cases of misuse.
