Open HaritzPuerto opened 3 years ago
Thanks for your interest. I think @czheng17 should be able to help you with the details of the code.
你好
首先,感谢您的工作。这真的很鼓舞人心!我读过你的论文,但我有几个问题,代码没有澄清它们。
- 如何创建初始节点嵌入?即,给定一个SRL参数,你如何构造你放入GCN的节点的嵌入?我在4.2中看到你使用GloVe作为图形编码器模块,但我不明白你如何创建嵌入。如果 SRL 有 3 个单词,如何聚合所有这 3 个单词的嵌入以创建节点嵌入?
- 从论文中:
但是,我在代码中找不到它。如何使用代码中的图形进行答案跨度预测?
在para_reader/模型中.py我发现了这个,但这里没有使用图表:
谢谢 Hello, did you solve it?
Hi,
First of all, thank you for your work. It is really inspiring! I have read your paper but I have a couple of questions and the code didn't clarify them.
1) How do you create the initial node embeddings? i.e., given an SRL argument, how do you construct the embedding of the node that you put into the GCN? I see in 4.2 that you use GloVe for the graph encoder module, but I don't see how you create the embedding. If the SRL has 3 words, how do you aggregate the embedding of all those 3 words to create the node embedding?
2) From the paper:![image](https://user-images.githubusercontent.com/2658728/97987316-6a86bf80-1e1e-11eb-8b02-0059ba82c22e.png)
However, I cannot find this in the code. How do you use the graph in the code for the answer span prediction?
In para_reader/model.py I found this, but the graph is not used here:![image](https://user-images.githubusercontent.com/2658728/97987477-b8032c80-1e1e-11eb-9f3f-066d94eca012.png)
Thank you