In this paper, we propose a method for estimating latent aspects in online review documents. Review aspects represent features of items or services evaluated by users. We can expect to acquire useful features for users by discovering their review aspects. We apply topic models to this problem. Existing work proposed methods for estimating the topics of the whole document or sets of sentences with various window sizes. In this paper, we propose two-level learning approach that connects adjacent sentences when their topics are similar, and re-estimates topics once again using the determined processing units. In the experiments of precision using perplexity, we confirm our proposed method improves on the existing method.