DeepMoji

Why we included this project

DeepMoji is a well-known model for sentiment and emotion analysis, trained on 1.2 billion tweets that include emojis. That training lets it pick up on tone and sarcasm in short informal text, and you can use it to score texts for emoji likelihood or to get 2304-dimensional emotional feature vectors for other tasks. The main appeal is transfer learning: you start from the pretrained weights and fine-tune on your own data instead of building a model from scratch, which is a big time-saver when you have limited labeled examples. The downside is the codebase is tied to Keras with Theano or TensorFlow backends and Python 2.7, so expect some porting work on modern stacks; if you use PyTorch, the torchMoji port is a smoother alternative. As a reference implementation for emotion-aware NLP, it's still worth studying, even if you end up adapting it heavily.