Online reviews as a new textual domain offer a unique proposition for sentiment analysis. Their short document length suggests any sentiment they contain is compact and explicit. Although supersized methods have obtained good results, a large amount of corpus should be trained beforehand. Recently, topic models have been introduced for the simultaneous analysis for sentiment in the document. However, the LDA model makes the assumption that, given the parameters the words in the document are all independent. It obviously isnt the case. The words in the document express the sentiment of the author. This paper proposes a model to solve the problem. We assume that the sentiments are related to the topic in the documents. A sentiment layer is added to the LDA model to improve it. Experimental result in the dataset demonstrates the advantage of the proposed model.