Abstract: Online comment is an effective way for users to express their opinions or suggestions on commodities．Analysis of comments is the basis of developing personalized services and improving the performance of commodities． However，there are a lot of false comments in the network． Accurate identification of the key comments which represent users' real feelings has become a hot issue in the field of comment analysis． A key comment extraction method based on reply support was proposed． The proposed method starts from user's feedback behavior to comments，and focuses on two indicators: like comment and comment reply． By calculating the like comment rate and comment reply rate，the method obtains the comment's reply support，only extracts the comment with high reply support，thus eliminating a large number of false or useless comments，and improves the accuracy of key comment extraction． Finally，through the experimental comparison with the existing mainstream methods，it is verified that the proposed method has a high accuracy and recall rate．

Key words: online comment, reply support, key sentences extraction, latent Dirichlet allocation ( LDA)