This document presents an unsupervised methodology for automatically learning social networks from multiple news sources. It uses a small number of seed syntactic templates to identify relationships between entities. These templates are then used to learn new syntactic patterns from news clusters expressing the same relationships. An efficient graph matching algorithm is then used to extract related entities and build the social network. The method achieves a precision of 0.61 and recall of 0.56 for identifying meeting relationships, and 0.57 precision and 0.10 recall for identifying support relationships.