Tech
"How it works part"
The whole prediction process is done on your browser via TensorflowJS (yes your browser is running AI!). Model is trained on a dataset scrapped from Shopee Reviews and hand labelled.
Sentences are first sanitized, common words are auto-corrected based on a common dictionary. Special characters and emojis are stripped and words are lowercase.
01101100 01101001 01100110 01100101 01101100 01100101 01110011 01110011
It is then converted to numbers (tokenized) for the algorithm to understand.
01100111 01101111 00100000 01100001 01110111 01100001 01111001
The tokens are then passed into a word2vec algorithm that returns a classification result in percentage [0.01, 0.89].
First value indicates it is 0.01 a good review, second indicates it is 0.89/89% a bad review.