search

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 the format of dataset #7

Closed chrislouis0106 closed 4 years ago

chrislouis0106 commented 4 years ago

hi， by reproducing the project of “Multihop with reward shaping”， I found that your datasets in MetaKGR are almost similar to Multihop-RS. At the same time, in the Multihop-RS, the NELL-995 dataset is different from you's. Meanwhile, by the project "DeepPath", I saw the other NELL-995 dataset. So, I'm confused with the format of the dataset. And I tried to find the original dataset of NELL-995. But I failed. Another a small question, I don't sure that I think that "raw.kb = train.triple + dev.triple + test.triple” which is right or not. If you could help me, I will appreciate it. Best wishes

davidlvxin commented 4 years ago

The NELL dataset in our work is derived from Multihop-RS. I did some processing on NELL-995 (from Multihop-RS) because the test set of the original NELL-995 data set only contains a small number of relations. In our work, we treat NELL-995 from Multihop-RS as the original dataset of NELL-995.

For the second question, raw.kb is equal to train.triple in our work.

chrislouis0106 commented 4 years ago

Thank you very much. the dataset in your project is easy to understand. But I don't sure how you get it by the NELL-995 in the Multihop-RS and could you explain it detailedly. image image another little problem, in my opinion, I feel that the raw is the original NELL-995 dataset. But I found that the size of raw.kb in the Multihop-RS is different from the size of raw.kb in MetakGR project. I don't know why and I don't found the original NELL-995 dataset.

davidlvxin commented 4 years ago

I randomly split the raw.kb in Multihop-RS and generate train.triples/dev.triples/test.triples. There is no problem with this processing, but there are some differences with the original NELL-995 in Multihop-RS. You can refer to README in here to understand the composition of the original NELL-995 in Multihop-RS.

chrislouis0106 commented 4 years ago

Enn, I got it. And I sum the length of files "train, dev, test" and found that the sum length is not equal to the length of raw.kb in the Multihop-RS. I don't know why. Haa.

chrislouis0106 commented 4 years ago

en, if you continue the work, I would like to add your WeChat and discuss with you about your project and Multihop project. Best wished!