In an active e-commerce or search environment, Natural Language Processing (NLP) and Text Mining can be the essential tasks, while the text data often co-exist with abundant and heterogeneous user behavior information in the same ecosystem. By leveraging large scale text plus user behavior data, we proposed a number of algorithms/applications to address different novel/classical research questions. In this talk, three cases will be presented: 1. Educate Machine to be a Creative Writer via User Feedback, 2. Anti-spam with Text and User BerryPicking Data, and 3. Sentiment Analysis via Heterogeneous Behavior-Graph Mining. In these works, the user behavior and text information are jointly encapsulated by using innovative representation learning, and the proposed models are able to bridge various kinds of data for efficient optimization.