Abstract: There are two main challenges for social media search. First, the messages are short, and are quite hard to construct indices. Second, the ranking list is too simple to fairly express the global structure of social media data. This paper proposes a visual search system which discovers inner semantic features from the raw data to strengthen index structures and provides an interactive interface to visualize features and filters search results. By using Twitter data as an example, our approach extracts topics and named entities based on temporal relationships. Then, a hierarchical semantic graph model is built to simplify semantic relations between topics and named entities. In the meantime, the model provides an essential index for visual query. During exploration, a set of split rings is employed to show multiple semantic patterns, together with informative interactions. Case studies demonstrate that the split ring representation preserves more obvious patterns than linked lines and bubble sets, and facilitates convenient search of messages of interest. User studies indicate that this system can find targets more easily than conventional systems.