|
--- |
|
language: |
|
- zh |
|
tags: |
|
- bert-base-chinese |
|
license: afl-3.0 |
|
--- |
|
This project page is about the pytorch code implementation of GlyphBERT by the HITsz-TMG research group. |
|
|
|
 |
|
|
|
|
|
GlyphBERT is a Chinese pre-training model that includes Chinese character glyph features.It renders the input characters into images and designs them in the form of multi-channel location feature maps, and designs a two-layer residual convolutional neural network module to extract the image features of the characters for training. |
|
|
|
|
|
The experimental results show that the performance of the pre-training model can be well improved by fusing the features of Chinese glyphs. GlyphBERT is much better than BERT in multiple downstream tasks, and has strong transferability. |
|
|
|
|
|
For more details about using it, you can check the [github repo](https://github.com/HITsz-TMG/GlyphBERT) |
