Extracting the social relation network of persons is challenging. Discovering significant binary person relations embedded in the web news would be appropriate as the starting point. Prior methods for this task, however, chose to define the relation types first, focused on a few limited types, and always took over a large amount of web information. This paper describes an unsupervised person relation extraction system. This system automatically extracts important people relations from a limited batch of web news, and then proceeds to cluster the instances of these relations and finds discriminative words to represent different clusters. We use various feature ranking strategies for filtering instead of simple bag-of-words representation. We present the experiments evaluation results and give an overview of possible enhancements of this system.