I have a question about the PubMed datasets

[jbjeong91]

Hi, I have a question about the PubMed datasets.

For example, there is a sentence, "Preclinical data have demonstrated that afatinib is a potent irreversible inhibitor of EGFR receptors including the T790M variant. And .... it is the T790M." and in this sentence, there are two relations, (resistance_nonrespose, afatinib, egfr, t790M) and (resistance_nonrespose, afatinib, egfr, t790M).

In this situation, the two instances are only used for the training(or test)? or is it possible that one is used for the training, the other one is used for the test?

Actually, in your research, if there are two relations in a sentence, the relations are used as individual instances. Thus, I think it is possible that one is used for training and the other one is used for the test. is it correct?

[VioletPeng]

We made it clear in the paper that there won't be overlap between the train/dev/test. Not only on the sentence level, we made. sure the train/dev/test data contains different PubMed articles.