Affiliations: [a] Department of Systems Innovation, The University of Tokyo, 7-3-1, Hongo, Bunkyo-ku, Tokyo, Japan. kenchin110100@gmail.com | [b] Department of Systems Innovation, The University of Tokyo, 7-3-1, Hongo, Bunkyo-ku, Tokyo, Japan. ohsawa@sys.t.u-tokyo.ac.jp
Correspondence:
[*]
Address for correspondence: Department of Systems Innovation, The University of Tokyo, 7-3-1, Hongo, Bunkyo-ku, Tokyo, Japan.
Abstract: It is useful to extract review sentences based on an assigned viewpoint for purposes such as summarization tasks. Previous studies have considered review extraction using semi-supervised learning or association mining. However, we approach this task using a clustering method. In particular, we focus on a topic model as a clustering method. In the conventional topic model, after randomly initializing the word distribution and the topic distribution, these distributions are estimated in order to minimize the perplexity using Gibbs sampling or variational Bayes. We introduce a new method called the PageRank topic model (PRTM) for estimating multinomial distributions over topics and words using network structure analysis methods. PRTM extracts topics by focusing on the co-occurrence relationships of words and it does not need randomly initialized values. Therefore, it can calculate unique word and topic distributions. In experiments using synthetic data, we showed that PRTM can infer an appropriate number of topics by clustering short sentences, and it was particularly effective when the sentences were covered by a small number of topics. Furthermore, in a real-world review data experiment, we showed that PRTM performed better with a shorter runtime compared with other models that infer the number of topics.
Keywords: network analysis, review data, sentence clustering, topic model
