Word emphasis prediction is an important part of expressive prosody generation in modern Text-To-Speech (TTS) systems. We present a method for predicting emphasized words for expressive TTS, based on a Deep Neural Network (DNN). We show that the presented method outperforms machine learning methods based on hand-crafted features in terms of objective metrics such as precision and recall. Using a listening test, we further demonstrate that the contribution of the predicted emphasized words to the expressiveness of the synthesized speech is subjectively perceivable.