In this paper, we propose a method for identifying opinion holders and opinion targets from Romanian texts, using an unsupervised approach. The method discriminates the opinion bearing sentences first. Then, it applies a set of rules for holder and target identification. We propose a series of identification rules in addition to an initial set existing in literature. We have performed a series of evaluations using data from several domains. The results indicate a significant performance improvement when the enhanced set of rules is considered, as opposed to the initial set.