Sentiment140 dataset with 1.6 million tweets

Dataset from Kaggle.

In this case, only using a subset which contains 300k tweets(lacking of GPU time). Evaluating the performance by 60,000(20%) tweets and training the model on 240,000(80%) tweets and finally achieving 81% accuracy.

The final model is constructed by 2 layers of Bi-LSTM, followed by CNN. And the output is logist.