Why use depth encoding in AST reader but not in query?

[brando90]

Hi Authors!

First I want to thank you for your ongoing patience and openness to clarify things for me. I appreciate it!

My question is that I noticed that in the AST reader according to the paper you use the depth embedding but in the query (= Non-Terminals) to the decoder you do not use the depth embedding but instead use the path embedding. I was trying to understand why you did not choose to use path embedding for both + depth embedding. Or even only use path embedding for both (since the path seems to encode the information the depth embedding would have inserted).

Is there a reason why the generated rules do not use the path embedding also like the query? That would have seemed like a more complete embedding since it includes the whole path and not only the parent.

Thanks for your time again and nice work.

[zysszy]

My idea is that the query used in our model is not non-terminals but a tree path from the root node to the node to be expanded, which contains the depth information implicitly.

However, we didn't conduct the experiment on path embedding + depth embedding or path embedding for both. These ideas may improve the performance of TreeGen.

Is there a reason why the generated rules do not use the path embedding also like the query? That would have seemed like a more complete embedding since it includes the whole path and not only the parent.

It is a good idea. We have not tried this idea before. It is due to that I think the tree convolutional layers + depth embedding is similar to path embedding.

Zeyu