Title: Twitter Sentiment Analysis via Bi-sense Emoji Embedding and Attention-based LSTM
Authors: Yuxiao Chen, Jianbo Yuan, Quanzeng You, Jiebo Luo
Publication: ACM Multimedia Conference 2018
1. Background
Previous work shows that, it is useful to pre-train a DNN on an emoji prediction task with pre-trained emoji embeddings to learn the emotional signals of emojis.
2. Problem
The previous emoji embedding models fail to handle the situation when the semantics or sentiments of the learned emoji embeddings contradict the information from the corresponding contexts.
Because they can only extract single embedding of each emoji.
3. Solution
Attention-based RNN (Recurrent Neural Network) with bi-sense emoji embeddings
Each emoji is embedded into 2 distinct vectors (positive-sense vector, negative-sense vector)
4. Dataset
Most of the previous Twitter sentiment datasets exclude emojis
and there exists little resource that contains sufficient emoji-tweets with sentiment labels.
So, the authors construct their own emoji-tweets datasets by automatically generating weak labels using a rule-based sentiment analysis algorithm Vader for pre-training the networks, and manually labeling a subset of tweets for fine-tuning and testing purposes.
5. Result
Bi-sense emoji embedding outperforms the SOTA
