Clinical Named Entity Recognition (CNER), the task of identifying the entity boundaries in clinical texts, is essential for many applications. Previous methods usually follow the traditional NER methods that heavily rely on language specific features (i.e. linguistics and lexicons) and high quality annotated data. However, due to the problem of Limited Availability of Annotated Data and Informal Clinical Texts, CNER becomes more challenging. In this paper, we propose a novel method that learn multiple representations for each category, namely category-multi-representation (CMR) that captures the semantic relatedness between words and clinical categories from different perspectives. CMR is learned based on a large scale unannotated corpus and a small set of annotated data, which greatly alleviates the burden of human effort. Instead of the language specific features, our proposed method uses more evidential features without any additional NLP tools, and enjoys a lightweight adaption among languages. We conduct a series of experiments to verify our new CMR features can further improve the performance of NER significantly without leveraging any external lexicons.