请问p71_TextRCNN_model.py中rnn-cnn layer定义原理是什么？最后ensemble left, embedding, right to output的方式不像Bi-LSTM layer呀

[JepsonWong]

请问p71_TextRCNN_model.py中rnn-cnn layer定义原理是什么？最后ensemble left, embedding, right to output的方式不像Bi-LSTM layer呀

[brightmart]

this is just like taking context information(from left and right) into considering when encode a position to get rich information.

…

________________________________ 发件人: JepsonWong <[email protected]> 发送时间: 2018年5月8日 10:35 收件人: brightmart/text_classification 抄送: Subscribed 主题: [brightmart/text_classification] 请问p71_TextRCNN_model.py中rnn-cnn layer定义原理是什么？最后ensemble left, embedding, right to output的方式不像Bi-LSTM layer呀 (#51) ― You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub<#51>, or mute the thread<https://github.com/notifications/unsubscribe-auth/ASuYMBBs4EfV-B6kxlCdTJQoiUPb648Nks5twQRVgaJpZM4T17Mv>.

[fei161]

The meaning of the paper should be two layers of lstm followed by cnn layer features, why is there no lstm and cnn in your code?

[JepsonWong]

When you get context right, you reverse the input sequences, but you don't reverse the output sequences. I think the output seqences should be reversed.
By the way, do you implement a rnn cell by yourself in your source code? So you don't use the tensorflow's api.
thanks.@brightmart

[JepsonWong]

I guess the author implement a rnn cell in his source code.

text_classification/a04_TextRCNN/p71_TextRCNN_model.py

Line 51 in a552e11

def instantiate_weights(self):

this is the rnn weights.
@fei161

[fei161]

The author first makes a full connection to the context. Then concatenates the context vector concat and maximizes it. Finally connects to the full connection layer.
@JepsonWong

[JepsonWong]

yes, i know this. I just think that when we get context right, we should reverse the output sequences.
@fei161