For a given comment on anything included in the train corpus, the model predicts its polarity (positive/negative) with an architecture that first converts the words into embedding vectors, then passes them through a neural network combining LSTM and convolutional layers to achieve better performance.
We collected and annoted a dataset containing 34639 Chinese comments and 13385 English comments on movies, books, music and electronic devices, which can be found in the dataset folder.
-
Chinese: 87.47%
-
English: 83.69%
-
You may use your own word embedding and train the model with a new language
-
The cake is a lie.
Negative -
I'm lovin' it.
Positive