With the development of online shopping, the Web has produced a large quantity of product reviews containing abundant evaluation knowledge about products. How to extract aspect and opinion words from the reviews and further obtain the sentiment polarity of the products at aspect level is the key problems to solve in fine-grained sentiment analysis of product reviews. First, considering certain features of Chinese product reviews, this paper designs methods to derive semantic relationships among words through syntactic analysis, word meaning understanding and context relevance, and then embed them as constrained knowledge into the topic model. Second, a semantic relation constrained topic model called SRC-LDA is proposed to guide the LDA to extract fine-grained topical words. Through the improvement of semantic comprehension and recognition ability of topical words in standard LDA, the proposed model can increase the words correlation under the same topic and the discrimination under the different topics, thus revealing more fine-grained aspect words, opinion words and their semantic associations. The experimental results show that SRC-LDA is an effective approach for fine-grained aspects and opinion words extraction.