THU-KEG / MetaKGR

Source codes and datasets for EMNLP 2019 paper "Adapting Meta Knowledge Graph Information for Multi-Hop Reasoning over Few-Shot Relations"
115 stars 21 forks source link

about code/data #2

Closed chuxuzhang closed 4 years ago

chuxuzhang commented 4 years ago

It seems that test data in github is not consistent with that in the paper. Could you provide the train/test data which can reproduce the results reported in Table 1? It would be great if you can provide some checkpoint information (e.g., how many epochs in training and adaptation) to reproduce the results in the paper. Thanks a lot.

davidlvxin commented 4 years ago

The test data in github is used for K=max setting, which is also used for Table 1.

Since the code and output files are not well organized, I can't find the exact checkpoint information. But I can still give the following information from my memory:

FB15K237: Training epoch: 15 or 25 Adaptation epoch: 39

NE: Training epoch: 40 Adaptation epoch: 40

Please feel free to contact me here if you have more questions.

chuxuzhang commented 4 years ago

Thanks for quick reply. The triple number in test data of FB in github is 668 while the test data of whole data in paper seems to be around 4k (raw data - training data). Could you double check it? Let me know if I made mistake.

davidlvxin commented 4 years ago

4k is not the number of test data, but the numbet of few-shot triples in train data.