dinobby/zs-bert

Official implementation of the paper "Towards Zero-Shot Relation Extraction with Attribute Representation Learning."

zs-bert's Issues

Where is the dataset ?

The link is not found now.

seed for the data splits

In the paper, you mentioned that you have run five times for different data split. what's the seed for your split?

Dataset links not working

Thank you for your work!
Currently, the google drive links for the datasets requires access, could you change the sharing permissions?

https://drive.google.com/file/d/1Ju5ngSL9Bpb7K0br11ocyPWzDA8N5-9v/view?usp=sharing
https://drive.google.com/file/d/1q_q1TA-c_xSCj0-CscWxmvvExr46ZcQS/view?usp=sharing

F1-score

the f1_score reported in your paper is micro_f1_score or macro_f1_score?

Missing License

Hi @dinobby ,
Thanks for contributing the source code of a very interesting work.

We are considering using you code for experiments and integration, but license information is not available.

Would it be possible to add a license to the repository?

The number of relations in FewRel dataset

Hi @dinobby ,

Thank you for releasing the source code of your nice work.
Sorry for disturbing you. However, I only wonder the number of relations of FewRel dataset that you reported in your paper.
In the original paper, the authors reported that the number of relations for FewRel is 100 relations and 70,000 instances. However, in Table 1 in your paper, you reported the number of relations is 80, with 56,000 instances.
Please let me know why they are different. Thank you very much for your time and your help!

No code...

fewshot result

about Wiki-ZSL dataset

Hi @dinobby ,
Thank you for releasing the source code of your nice work.
Sorry for disturbing you. In the Wiki-ZSL data, each original sample contains multiple relations, that is, the length of 'edgeSet' is greater than 1. And in the code of processing Wiki-ZSL data, only the first item of 'edgeSet' is selected, that is, only one relation category is considered for each sample, and other categories are directly ignored. Am I understanding that right? Why not consider all relations for each sample?
Hope to get your answer and thank you very much for your time and your help!