We present a knowledge-based system to extract product feature-orientation (sentiment) pairs from on-line product reviews. Unlike the vast majority of existing approaches, our system first extracts strong implicit opinions, before searching for explicit product feature keywords. We call this the "opinion (O) first, feature (F) second" approach, which incidentally seems to work well with Chinese reviews. Our system relies heavily on a hierarchical product feature concept model (ontology) that lists popular feature and opinion vocabulary pertaining to a product genre. The concept model is built manually using product domain knowledge and subsequently expanded via a Chinese semantic lexicon. To the best of our knowledge, our work is among one of the first studies on Chinese product feature review extraction at the sentence segment resolution. Experiments comparing our approach to a well-known review mining algorithm shows the feasibility and robustness of our system.