Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Arabic T5v1.1 for question paraphrasing
|
| 2 |
+
|
| 3 |
+
This is a fine-tuned [arabic-t5-small](https://huggingface.co/flax-community/arabic-t5-small) on the task of question paraphrasing.
|
| 4 |
+
|
| 5 |
+
A demo of the trained model using HF Spaces can be found [here](https://huggingface.co/spaces/salti/arabic-question-paraphrasing)
|
| 6 |
+
|
| 7 |
+
## Training data
|
| 8 |
+
|
| 9 |
+
The model was fine-tuned using the [Semantic Question Similarity in Arabic](https://www.kaggle.com/c/nsurl-2019-task8/data) data on kaggle.
|
| 10 |
+
|
| 11 |
+
Only the rows of the dataset where the label is `True` (the two questions have the same meaning) were taken.
|
| 12 |
+
|
| 13 |
+
The training data was then also mirrored; so if `q1` and `q2` were two questions with the same meaning, then `(q1, q2)` and `(q2, q1)` were both present in the training set. The evaluation set was kept unmirrored of course.
|
| 14 |
+
|
| 15 |
+
## Training config
|
| 16 |
+
|
| 17 |
+
| | |
|
| 18 |
+
| :-------------: | :------: |
|
| 19 |
+
| `batch size` | 128 |
|
| 20 |
+
| `dropout rate` | 0.1 |
|
| 21 |
+
| `learning rate` | 0.001 |
|
| 22 |
+
| `lr schedule` | constant |
|
| 23 |
+
| `weight decay` | 1e-7 |
|
| 24 |
+
| `epochs` | 3 |
|
| 25 |
+
|
| 26 |
+
## Results
|
| 27 |
+
|
| 28 |
+
| | |
|
| 29 |
+
| :---------------: | :----: |
|
| 30 |
+
| `training loss` | 0.7086 |
|
| 31 |
+
| `evaluation loss` | 0.9819 |
|
| 32 |
+
| `meteor` | 49.277 |
|
| 33 |
+
| `sacreBLEU-1` | 57.088 |
|
| 34 |
+
| `sacreBLEU-2` | 39.846 |
|
| 35 |
+
| `sacreBLEU-3` | 29.444 |
|
| 36 |
+
| `sacreBLEU-4` | 22.601 |
|
| 37 |
+
| `Rouge F1 max` | 1.299 |
|