Summarization of online reviews by customers is a popular practice for evaluation of products or services. As the reviews accumulate, the large size and the unstructured nature of the reviews hinder manual summarization. Automatic categorization of the reviews as a whole into only positive and negative group cannot represent a clear picture. An aspect based automatic summarization technique can provide better visualization. However, automatic extraction of proper aspects from the huge reviews of any product is not very easy. There are some research works in this direction, but any definite method is yet to come. In this work, a two-step Latent Dirichlet Allocation (LDA) technique, which is popularly used for topic modelling has been developed for efficient aspect extraction. The method has been evaluated by simulation experiments on Amazon product reviews and Yelp restaurant and hotel reviews. The results have been found quite matching with human annotated results.