Many natural language applications need to represent the input text into a fixed-length vector. Existing technologies such as word embeddings and document representation provide natural representation for natural language tasks, but they do not consider the importance of each word in the sentence, and also ignore the significance of a sentence in a document. This study proposes a Document Representation model based on a Hierarchical Attention (HADR) mechanism, taking into account important sentences in document and important words in sentence. Experimental results show that documents that take into account the importance of words and importance of sentences have better performance. The accuracy of this model in the sentiment classification of documents (IMBD) is higher than that of Doc2Vec and Word2Vec models.