The recent development of social media (e.g., Twitter, Facebook, blogs, etc.) provides an unprecedented opportunity to study human social cultural behaviors. These data sources provide rich structured data (e.g., XML, relational tables, and categorical data) as well as unstructured data (e.g., texts). A significant challenge is to summarize and navigate structured data together with unstructured text data for efficient query and analysis. In this paper we introduce a text cube architecture designed to organize social media data in multiple dimensions and hierarchies for efficient information query and visualization from multiple perspectives. For example, an affective process cube allows the analyst to examine public reaction (e.g., sadness, anger) to a range of social phenomena. The text cube architecture also supports the development of prediction models using the summarized statistics stored in a data cube. For example, models that detect events, such as violent protests in the Egyptian Revolution, can be built using the linguistic features stored in an event data cube. These kinds of models represent higher level of knowledge representation and may help to develop more effective strategies for decision-making based on social media data.