Techniques are provided for performing sentiment analysis on words in a first data set. An example embodiment includes generating a word embedding model including a first plurality of features. A value indicating sentiment for the words in the first data set can be determined using a convolutional neural network (CNN). A second plurality of features are generated based on bigrams identified in the data set. The bigrams can be generated using a co-occurrence graph. The model is updated to include the second plurality of features, and sentiment analysis can be performed on a second data set using the updated model.