A new framework for building Natural Language Processing (NLP) models in the Deep Learning era has been proposed by Matthew Honnibal (creator of the SpaCy NLP toolkit). It is composed of the following four steps - Embed, Encode, Attend and Predict. Embed converts incoming text into dense word vectors that encode its meaning as well as its context; Encode adapts the vector to the target task; Attend forces the network to focus on the most important parts of the data; and Predict produces the network's output representation. Word Embeddings have revolutionized many NLP tasks, and today it is the most effective way of representing text as vectors. Combined with the other three steps, this framework provides a principled way to make predictions starting from unstructured text data. This presentation will demonstrate the use of this four step framework to build Deep Neural Networks that do document classification and predict similarity between sentence and document pairs, using the Keras Deep Learning Library for Python.