HAN은 구조도 더 복잡하고, ATTENTION 까지사용, 그리고 GLOVE를 이용해 단어임베딩도 사용했는데도 95 그냥 LSTM은 임베딩부터 학습하고 ATTENTION도 안썼는데, 그냥 문단을 LSTM 사용하니 성능이 99

같은 데이타인데 왜그럴까?

HAN

An aexample of HAN from paper: http://www.cs.cmu.edu/~./hovy/papers/16HLT-hierarchical-attention-networks.pdf. , NAACL-HLT, 16'

Code was combined from two publickly available sources: https://gist.github.com/cbaziotis/7ef97ccf71cbc14366835198c09809d2

https://github.com/richliao/textClassifier

https://richliao.github.io/supervised/classification/2016/12/26/textclassifier-HATN/

https://hist0134.blog.me/221179965199

https://m.blog.naver.com/PostView.nhn?blogId=hist0134&logNo=221386940063&navType=tl

설명 : http://hugrypiggykim.com/2018/05/31/han-hierarchical-attention-networks-for-document-classification/

1. data 폴더에 labeledTrainData.tsv 필요
   • wget https://www.kaggle.com/c/word2vec-nlp-tutorial/download/labeledTrainData.tsv
   • 이름바꾸고 unzip
2. glove 폴더에 glove.6B.100d or glove.6B.200d, (단어, 임베딩)쌍으로
   • http://nlp.stanford.edu/data/glove.6B.zip