We investigate systems that identify opinion expressions and assigns polarities to the extracted expressions. In particular, we demonstrate the benefit of integrating opinion extraction and polarity classification into a joint model using features reflecting the global polarity structure. The model is trained using large-margin structured prediction methods.