search

studio-ousia / luke

LUKE -- Language Understanding with Knowledge-based Embeddings
Apache License 2.0
705 stars 102 forks source link

Entity Linking Result？ #63

Closed dalek-who closed 3 years ago

dalek-who commented 3 years ago

Hi authors, I notice that

Global Entity Disambiguation with Pretrained Contextualized Embeddings of Words and Entities https://arxiv.org/abs/1909.00426

is also your work, and that one is nearly the same as LUKE except several slight difference. It is also awesome and is Entity Linking SOTA, but why don't you show the Entity Linking perfromance in LUKE?

ikuyamada commented 3 years ago

Hi, Thank you for raising the issue. The code used in the entity linking paper is also based on this repository. We plan to release the trained checkpoints in the future.

dalek-who commented 3 years ago

Thanks, waiting for your checkpoint. Can current checkpoint reproduce that linking result? I want to try it in my system. Also, I have two question about that entity linking paper:

  1. In the "5 Experiment" chapter's second paragraph, We mask 90% of the mentions and fix the entity token embeddings (B and B ∗ ) and the output bias (b oand b o∗ ).，how to understand mask 90% of the mentions? Why not use all mentions to fineturn entity link?
  2. That entity linking pretrain takes 8 V100 10 days, but LUKE takes 16 V100 30days, why?
ikuyamada commented 3 years ago

We did not test the performance of entity linking using the current checkpoint.

In the "5 Experiment" chapter's second paragraph, We mask 90% of the mentions and fix the entity token embeddings (B and B ∗ ) and the output bias (b oand b o∗ ).，how to understand mask 90% of the mentions? Why not use all mentions to fineturn entity link?

Since the paper uses global context based on the entity sequence to disambiguate mentions, it is required to train the model to use the non-masked entities as contexts. If we mask 100% of mentions, the model cannot learn such interactions during fine-tuning.

That entity linking pretrain takes 8 V100 10 days, but LUKE takes 16 V100 30days, why?

We used different servers with different training steps and batch size :) Please note that this repository is not the official code of the entity linking paper.

dalek-who commented 3 years ago

Thanks!

ikuyamada commented 2 years ago

Hi @dalek-who, I've just released the code and model checkpoint file here.

dalek-who commented 2 years ago

Thank you for this excellent work !!

在 2021年11月25日，下午2:33，Ikuya Yamada @.***> 写道：

Hi @dalek-who https://github.com/dalek-who, I've just released the code and model checkpoint file here https://github.com/studio-ousia/luke/tree/master/examples/entity_disambiguation.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/studio-ousia/luke/issues/63#issuecomment-978878587, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHT27AWHRKWKW34JZGH6KZLUNXKDDANCNFSM42KKKDIA. Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.