Hierarchical Attention Network is a network architecture that leverages attention mechanism to build a context vector for a document which is finally passed to a softmax function in a classification problem. It uses attention in a hierarchical fashion. Words form a sentence and sentences form a document. Both at word level and sentence level, it uses attention mechanism to give importance to relevant words in a sentence as well as relevant sentences in a document. 

Hierarchical Attention Networks give great results in text classification tasks. Original paper ( Hierarchical Attention Networks for Document Classification ) gives SOTA results on various public datasets for text classification