alibaba-nlp / muver Goto Github PK

View Code? Open in Web Editor NEW

30.0 30.0 0.0 25 KB

[EMNLP 2021] MuVER: Improving First-Stage Entity Retrieval with Multi-View Entity Representations

Python 100.00%

entity-linking entity-retrieval

muver's People

Contributors

Stargazers

Watchers

muver's Issues

一个bug

你好！

MuVER/muver/multi_view/train.py

Line 221 in 28761d9

if args.do_train:

这一行，应该改为：
if args.do_train and args.local_rank in [0, -1]:

按照代码逻辑，local rank不是0/1的进程 log==None，对于他们来说并没有checkpoint_path这个参数

关于 [train/valid/test]_top_64_candidates.tsv 文件

你好！
我想请问一下，

MuVER/muver/multi_view/data_loader.py

Line 283 in d1da6eb

    
           candidate_path = os.path.join(candidate_path, "{}_top_64_candidates.tsv".format(mode))

这一行的 *_top_64_candidates.tsv 这个是自己处理得到的？还是zeshel原始论文里面说的使用tf-idf检索得到的64个候选？

how long to test the model without muti view？why so long for me to exec it in v100（32g）？

What --pretrained_model should I use?

Hi, thanks for your contribution. When I tried to reproduce the result by using provided checkpoint, --pretrained_model arguments indicates path_to_model/bert-base, but I don't know which model should I use ( what is 'path_to_model' you used in training?). Can you help me? Thank you!

Recommend Projects