|
--- |
|
language: |
|
- multilingual |
|
tags: |
|
- hack |
|
datasets: |
|
- Wikipedia |
|
--- |
|
|
|
|
|
## bert-base-multilingual-cased-segment1 |
|
|
|
This is a version of multilingual bert (bert-base-multilingual-cased), where the segment embedding of the 1's is copied into the 0's. Yes, that's all there is to it. We have found that this improves performance substantially in low-resource setups for word-level tasks (e.g. average 2.5 LAS on a variety of UD treebanks). More details are to be released in our LREC2022 paper titled: Frustratingly Easy Performance Improvements for Cross-lingual Transfer: A Tale on BERT and Segment Embeddings. |
|
|
|
These embeddings are generated by the following code |
|
``` |
|
import AutoModel |
|
baseEmbeddings = AutoModel.from_pretrained("bert-base-multilingual-cased") |
|
tte = baseEmbeddings.embeddings.token_type_embeddings.weight.clone().detach() |
|
baseEmbeddings.embeddings.token_type_embeddings.weight[0,:] = tte[1,:] |
|
``` |
|
|
|
More details and other varieties can be found in the repo: https://bitbucket.org/robvanderg/segmentembeds/ |
|
|
|
Note that when using this model on a single sentence task (or word-level task), the results would be similar as just using `token_type_id=1` for all tokens. |
