Social media sites and blogs generate a vast amount of emotionally rich data in the form of tweets, status updates, blog posts etc. Such textual data are a good representative of emotions expressed by an individual or a group of people on any given topic. By analyzing the emotions within these textual data, we can get an idea about how an individual or a community expresses their views. Analytical techniques are widely used for analyzing emotions within these texts. However, due to the imbalanced nature of the training datasets the supervised classifiers fail to clearly classify the different emotion classes. As a result, these classifiers demonstrate a poor performance in identifying emotions within the texts. Here, using a constructed heterogeneous training dataset from well-known training datasets we have trained two deep learning models namely the Convolutional Neural Network (CNN), and Recurrent Neural Network (RNN) to address a four-class emotion (Anger, Sadness, Happy, Surprise) classification problem. By appropriately tuning the hyper parameters of the deep learning classifiers our study reveals that the CNN classifier has a slightly better performance (77%) than the RNN classifier (76%) for a four-class emotion classification problem.