CSGTM: Capsule Semantic Graph-Guided Latent Community Topics Discovery

Topic modeling plays an important role in text mining for semantic mining. Existing topic models remain semantic loss, fuzzy topic concepts, and topic overlap issues because the sparsity of semantic information. For addressing above issues, in this paper, we propose a novel topic discovery method based on graph analysis named CSGTM (Capsule Semantic Graph Topic Model). Firstly, we utilize the community detection algorithm ESLPA (Entropy Similarity Label Propagation Algorithm) proposed in this paper to decompose the corpus into different semantic units, and then build the relationship between semantic units to form a capsule semantic graph. Secondly, CSGTM infers the topic distribution of each capsule vertex within a narrow range from community-level to accomplish topic modeling by assuming that topics and communities belong to different concepts. Experimental results on four benchmark datasets in various domains verify the superiority and efficiency of our proposed method CSGTM.