Joint Emoji Classification and Embedding Learning

Under conversation scenarios, emoji is widely used to express humans' feelings, which greatly enriches the representation of plain text. Plentiful utterances with emoji are produced by humans manually in social media platforms every day, which make emoji great influence on the human life. For the academic community, researchers are always with the help of utterances including emoji as annotated data to work on sentiment analysis, yet lack of adequate attention to emoji itself. The challenges lie in how to discriminate so many different kinds of emoji, especially for those with similar meanings, which make this problem quite different from traditional sentiment analysis. In this paper, in order to gain an insight into emoji, we propose a matching architecture using deep neural networks to jointly learn emoji embeddings and make classification. In particular, we use a convolutional neural network to get the embedding of the utterance and match it with the embedding of the corresponding emoji, to obtain its best classification, and otherwise also train the emoji embeddings. Experiments based on a massive dataset demonstrate the effectiveness of our proposed approach better than traditional softmax methods in terms of p@1, p@5 and MRR evaluation metrics. Then a test of human experience shows the performance could meet the requirement of practice systems.