frankwork / acnn Goto Github PK

你好，复现代码里的这个senna embedding是怎么得到的啊？

I think the attention of the pooling layer tries to get the relations between the i-th word and j-th class, so the U should be (d^context(after conv), d^class_embedding), but the implementation of the U is (d^context, num_class).

And the loss part, I think the paper have a little flaw, or maybe I was wrong. When we try to separate the classes, we should try to distinguish the classes that are hard to split, so we should trying decrease the negative y. We need choose the argmin \delta(negative y).

Maybe I am wrong about that, looking forward the apply, thanks.

hello,I choose attention_pooling model when I run main.py. But I see a great difference from your results.And the train accuracy is so high, the test accuracy is so low.
My result in attention pooling model as follows:
Epoch: 1 Train: 28.94% Test: 42.63%
Epoch: 10 Train: 67.42% Test: 54.52%
Epoch: 50 Train: 92.29% Test: 55.44%
Epoch: 100 Train: 94.67% Test: 53.89%

Looking forward to your reply.

你这个是可以实现的版本吗？使用tensorflow?

Hi, my implementation is similar as yours. In input attention layer, I did convolution of kernel size 3 first and then multiply with the attention. I didn't see a mathematical difference between this version and the sliding window version. What's your opinion on it?

Hi,

In function pos in file utils.py, relative distance of words is mapped to [0,123). Why is 123 chosen? Is it related to the maximum sentence length of the data?

Thanks,
Nigel

I noticed that in the log file, the version "baseline+attentive pooling" can get the result: 05-10 21:12 Epoch: 21 Train: 94.81% Test: 75.19%. What are the model configurations in details for this result? If possible, could you send me the model file for this result? My email is [email protected]. I have tried my best but cannot reach this performance. Thank you so much!

please introduce simplely this dataset ,include the every number in front of the sentence