Automatic encoding of diagnosis and procedures can increase the interoperability and efficacy of the clinical cooperation. The concept, rule-based and machine learning classification methods for automatic code generation can easily reach their limit due to the handcrafted rules and a limited coverage of the vocabulary in a concept library. As the first step to apply deep learning methods in automatic encoding in the clinical domain, a suitable semantic representation should be generated. In this work, we will focus on the embedding mechanism and dimensional reduction method for text representation, which mitigate the sparseness of the data input in the clinical domain. Different methods such as word embedding and random projection will be evaluated based on logs of query-document matching.