hkust-knowcomp/bmgf-roberta

Source Code for IJCAI 2020 paper "On the Importance of Word and Sentence Representation Learning in Implicit Discourse Relation Classification"

bmgf-roberta's Issues

Questions about num_rels

parser.add_argument("--num_rels", type=int, default=4,
choices=[3, 4, 11, 12, 14],
help="how many relations are computed")

The choices of the hyper-parameter num_rels is [3, 4, 11, 12, 14], but in the conll dataset, the class is 15. So how to choose num_rels for the conll dataset?
By the way, could you please offer the whole hyper-parameters for the conll dataset? And, could you please describe the meaning of [3, 4, 11, 12, 14], respectively?

test数据集的使用

请问为何要在训练过程保存在test数据集上最优的模型，这样划分验证集和test集的意义是不是就没有了？这样得到的结果和别的模型作比较是否不公平？