link: https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge
get muse fasttext word embedding from https://github.com/facebookresearch/MUSE
get glove embedding from https://nlp.stanford.edu/projects/glove/
get crawl fasttext word embeding from https://www.kaggle.com/yekenot/pooled-gru-fasttext/data
-
train base NN models
-
train base lr, mnb and other models
-
combine with other features, use xgb or lgb to ensemble (10 fold, pub score 9871, pri score 9868)
-
blend with other results (pub score 9873, pri score 9869, final rank 99/45xx, top 3%， toxic小分队)