Hi,

(1) The data size in this paper is "the training set with 12,775 instances (Section 2-20), the validation set with 1,183 instances (Section 0-1), and the test set with 1,046 instances (Section 21-22)"， Is this before or after processing?

*Processing means: "Further, there exist 16 second-level labels, five of which with few training instances and no validation and test instance are removed. Therefore, we conduct an 11-way classification on the second-level labels. " in the paper.

(2) And in paper "On the Importance of Word and Sentence Representation Learning in Implicit Discourse Relation Classification", the data size is Train/dev/test: 12362, 1183, 1046. Which one is right and If there any code to process the dataset.

Thank you!

Hi, When will the code be released, Thanks?

Hi,

Thanks for your interesting work. I am confused about two parts of your implementation.

The first one is the data proprecessing. In the preprocess function, I see you maintain several "other" arrays, such as arg1_train_other and arg2_train_other and so on. If a sample doesn't have a second-level sence label, it will be assigned with a default label and added to those arrays. When evaluating, you calculate results based on the predictions on both normal samples and samples with default labels. This is fine for top-level evaluation, but not the case for second-level. Because previous works usually consider only samples with gold second level labels.

The second one is the construction of the graph. You mentioned in your paper that nodes will have a self-loop edges. But in the implementation, there is not self-loop in the adjacent matrix your provided. This contradicts your description.