Hierarchical Dirichlet process (HDP) is a non-parametric Bayesian model, and has been widely applied in the application of topic models. However, the model is based on the "bag of words" hypothesis, ignoring the order of words in the document, resulting in a lack of word context semantics. In this regard, this paper proposes a window-based hierarchical Dirichlet process model (WHDP). The model uses windows to divide documents into smaller fragments, and keeps the order between words while moving windows, so as to reduce the semantic confusion of the text. We applied our method in real dataset and compared with other existing methods, such as sampling belief propagation algorithm for HDP, LDA model, and slidingwindow based topic model. The results show that the proposed method performs the superiority in convergence rate, perplexity and generalization ability.
Paper
Full text
Window-Based Topic Model for HDP
Semantic Scholar · Computer Science · 2019
Abstract
Hierarchical Dirichlet process (HDP) is a non-parametric Bayesian model, and has been widely applied in the application of topic models. However, the model is based on the "bag of words" hypothesis, ignoring the order of words in the document, resulting in a lack of word context semantics. In this regard, this paper proposes a window-based hierarchical Dirichlet process model (WHDP). The model uses windows to divide documents into smaller fragments, and keeps the order between words while moving windows, so as to reduce the semantic confusion of the text. We applied our method in real dataset and compared with other existing methods, such as sampling belief propagation algorithm for HDP, LDA model, and slidingwindow based topic model. The results show that the proposed method performs the superiority in convergence rate, perplexity and generalization ability.