Opinionated expression extraction is a central problem in fine-grained sentiment analysis. Most existing works focus on either generic subjective expression or aspect expression extraction. However, in opinion mining, it is often desirable to mine the aspect specific opinion expressions (or aspectsentiment phrases) containing both the aspect and the opinion. This paper proposes a hybrid generative-discriminative framework for extracting such expressions. The hybrid model consists of (i) an unsupervised generative component for modeling the semantic coherence of terms (words/phrases) based on their collocations across different documents, and (ii) a supervised discriminative sequence modeling component for opinion phrase extraction. Experimental results using Amazon.com reviews demonstrate the effectiveness of the approach that significantly outperforms several state-of-the-art baselines.