Online reviews provide facility so that internet user can give review about an aspect. Sentiments about a product are useful and have an influence in decision-making by person or organization. As in an opinion, eviewers and provide positive and negative reviews simultaneously. This is due, opinions targets are often not the product as a whole, but rather part of a product called the feature, where there are advantages and disadvantages in the eyes of reviewers. In this research, sentiment will be identified based on its opinion. Opinion data used in this research is in English, taken from the site www.cnet.com. The product conclusions presented based on product features. Thus, there are two processes undertaken in this research: (1) Extraction of product features in opinion, (2) Sentiment identification for each product feature. Feature extraction is done by searching for phrases that match the relation dependencies template, and then do the filtering feature. In sentiment identification, the positive and negative probability value, and also the target class of the feature opinion, became S3VMs input parameters. In the study by S3VMs, some data are treated as unlabeled data. Results obtained from this study for the evaluation of sentiment identification with F1-Measure at 86% for positive class and 70% for negative class. As for feature identification obtained 82% accuracy. For further development of this research, Improve SVM is suggested to handle the unbalance data problem. Mapping to implicit feature is also advisable to identify more product feature.