Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,51 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language: ug
|
3 |
+
license: mit
|
4 |
+
---
|
5 |
+
|
6 |
+
# gpt2-wechsel-uyghur
|
7 |
+
|
8 |
+
Model trained with WECHSEL: Effective initialization of subword embeddings for cross-lingual transfer of monolingual language models.
|
9 |
+
|
10 |
+
See the code here: https://github.com/CPJKU/wechsel
|
11 |
+
|
12 |
+
And the paper here: https://arxiv.org/abs/2112.06598
|
13 |
+
|
14 |
+
## Performance
|
15 |
+
|
16 |
+
| Model | PPL |
|
17 |
+
|---|---|
|
18 |
+
| `gpt2-wechsel-sundanese` | **111.72** |
|
19 |
+
| `gpt2` (retrained from scratch) | 149.46 |
|
20 |
+
|
21 |
+
| Model | PPL |
|
22 |
+
|---|---|
|
23 |
+
| `gpt2-wechsel-scottish-gaelic` | **16.43** |
|
24 |
+
| `gpt2` (retrained from scratch) | 19.53 |
|
25 |
+
|
26 |
+
| Model | PPL |
|
27 |
+
|---|---|
|
28 |
+
| `gpt2-wechsel-uyghur` | **34.33** |
|
29 |
+
| `gpt2` (retrained from scratch) | 42.82 |
|
30 |
+
|
31 |
+
| Model | PPL |
|
32 |
+
|---|---|
|
33 |
+
| `gpt2-wechsel-malagasy` | **14.01** |
|
34 |
+
| `gpt2` (retrained from scratch) | 15.93 |
|
35 |
+
|
36 |
+
See our paper for details.
|
37 |
+
|
38 |
+
## Citation
|
39 |
+
|
40 |
+
Please cite WECHSEL as
|
41 |
+
|
42 |
+
```
|
43 |
+
@misc{minixhofer2021wechsel,
|
44 |
+
title={WECHSEL: Effective initialization of subword embeddings for cross-lingual transfer of monolingual language models},
|
45 |
+
author={Benjamin Minixhofer and Fabian Paischer and Navid Rekabsaz},
|
46 |
+
year={2021},
|
47 |
+
eprint={2112.06598},
|
48 |
+
archivePrefix={arXiv},
|
49 |
+
primaryClass={cs.CL}
|
50 |
+
}
|
51 |
+
```
|