In the paper Hierarchical Attention Networks for Document Classification (http://www.cs.cmu.edu/~./hovy/papers/16HLT-hierarchical-attention-networks.pdf), Yang et al. have used a hierarchical deep learning architecture for classifying documents. The main intuition of their idea is that documents have inherent hierarchical structures such as words form sentences and sentences form documents. Moreover, not all parts of a document are equally important in capturing the semantics and meaning required for answering a specific query. They utilize two types of attention, one at the word level and another at the sentence level. In addition to that, the text representation is also split into a sentence level and document level. The following diagram shows the different components of the network that are capturing the word level and sentence level deep representations and the corresponding attention mechanisms:

 A Keras implementation of the model presented in the paper is available in GitHub https://github.com/richliao/textClassifier and the TensorFlow implementation can be found under https://github.com/ematvey/hierarchical-attention-networks.