search

rikeda71 / Char-BLSTM-CRF-for-Japanese

An Inplementation of Character based Bidirectional LSTM CRF for Japanese
MIT License
5 stars 1 forks source link

figure1 cnn kernel size #7

Closed hitxujian closed 5 years ago

hitxujian commented 5 years ago

thanks for sharing! in figure 1 of paper,how many filters in cnn and each filter size is how

rikeda71 commented 5 years ago

@hitxujian I'm not the author of the paper that proposed this method. So, I'm sorry if the following answer is wrong.

This method don't use CNN. Author's proposed model is figure 2 in the paper. Misawa et al. indicated CNN that is unsuitable for extracting Japanese sub-word information.

By the way, figure 1 model is BiLSTM-CNN-CRF model proposed by Ma et al. (https://arxiv.org/abs/1603.01354)