Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

COLING-2020-Specializing Unsupervised Pretraining Models for Word-Level Semantic Similarity #358

Open
BrambleXu opened this issue Feb 1, 2023 · 0 comments
Assignees
Labels
BERT(M) BERT Model Embedding Embedding/Pre-train Model/Task SSim(T) Semantic Similarity

Comments

@BrambleXu
Copy link
Owner

BrambleXu commented Feb 1, 2023

Summary:

为了将word-level上的语义相似度融合进预训练里,本研究对BERT进行了扩展,除了原本的masked language modeling和next sentence prediction两个任务,还添加了一个binary word relation classification的任务。这个任务的目的就是为了学习到文本里不同单词之间的相似度

Resource:

  • pdf
  • code 基于TF,用起来够呛
  • [paper-with-code](

Paper information:

  • Author:
  • Dataset:
  • keywords:

Notes:

先用fastText学习得到词向量的固定表征,利用这个构建c=(w1, w2),w1和w2是一组同义词。c代表一个positive training example。然后利用词向量,创建两个negtive example,c1* = (w1*, w2), c2* = (w1, w2*)。w1*是group里除了w1外,最靠近w2的单词。这样可以得到2倍数量呃negtive examples。

为了方便BERT训练,用wordpiece对w1,w2分割成wordpiece tokens,然后把两个近义词 (mended, regenerated) 创建成下面形式。w1和w2的segment ID是0和1.

image

预训练的时候,添加了一个新的分类器,一个 binary classifier,用于判断上面的一组数据,是否是lexico-semantic relation相关的(整个数据量,是1倍的positve, 2倍的negtive)。

MLM和NSP使用的是同一份数据,LRC是另一份数据。

Model Graph:

Result:

在一些涉及到语义相似的任务上,LIBERT效果更好。该研究还设计了一个 Similarity-Oriented Downstream Evaluation: Lexical Simplification任务,用于验证LIBERT在word-level语义相似度上的效果。比BERT好了0.9个百分点。

image

image

Thoughts:

Next Reading:

@BrambleXu BrambleXu self-assigned this Feb 1, 2023
@BrambleXu BrambleXu added Embedding Embedding/Pre-train Model/Task BERT(M) BERT Model SSim(T) Semantic Similarity labels Feb 1, 2023
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
BERT(M) BERT Model Embedding Embedding/Pre-train Model/Task SSim(T) Semantic Similarity
Projects
None yet
Development

No branches or pull requests

1 participant