Within multi-domain sentiment analysis, we study how different domain-dependent polarities can be learned for the same concepts. To this aim, we extend an existing approach based on the propagation of fuzzy polarities over a semantic graph capturing background linguistic knowledge to learn concept polarities with respect to various domains and their uncertainty from labeled datasets. In particular, we use POS tagging to refine the association between terms and concepts and word embedding to enhance the construction of the semantic graph. The proposed approach is then evaluated on a standard benchmark, showing that the combined use of POS tagging and word embedding improves its performance. One particularly strong point of the proposed approach is its recall, which is always very close to 100%. In addition, we observe that it exhibits good cross-domain generalization capabilities.