Topic Discovery based on LDA_col Model and Topic Significance Re-ranking

This paper presents a method to find the topics efficiently by the combination of topic discovery and topic re-ranking. Most topic models rely on the bag-of-words(BOW) assumption. Our approach allows an extension of LDA model—Latent Dirichlet Allocation_Collocation (LDA_col) to work in corpus such that the word order can be taken into consideration for phrase discovery, and slightly modify the modal for modal consistency and effectiveness. However, LDA_col results may not be ideal for user’s understanding. In order to improve the topic modeling results, two topic significance re-ranking methods (Topic Coverage(TC) and Topic Similarity(TS)) are proposed. We conduct our method on both English and Chinese corpus, the experimental results show that themodified LDA_col discovers more meaningful phrases and more understandable topics than LDA and LDA_col.Meanwhile, topic re-ranking method based on TC performs better than TS, and has the ability of re-ranking the “significant” topics higher than “insignificant” ones.