5roop commited on
Commit
1c9266e
1 Parent(s): 7e58a1f

Created model card

Browse files
Files changed (1) hide show
  1. README.md +49 -0
README.md ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # bcms-bertic-frenk-hate
2
+
3
+ Text classification model based on `classla/bcms-bertic` and fine-tuned on the [FRANK dataset](https://www.clarin.si/repository/xmlui/handle/11356/1433) comprising of LGBT and migrant hatespeech. Only the Croatian subset of the data was used for fine-tuning and the dataset has been relabeled for binary classification (offensive or acceptable).
4
+
5
+ ## Fine-tuning hyperparameters
6
+
7
+ Fine-tuning was performed with `simpletransformers`. Beforehand a brief hyperparameter optimisation was performed and the presumed optimal hyperparameters are:
8
+
9
+ ```python
10
+
11
+ model_args = {
12
+ "num_train_epochs": 12,
13
+ "learning_rate": 1e-5,
14
+ "train_batch_size": 74}
15
+ ```
16
+
17
+ ## Performance
18
+
19
+ The same pipeline was run with two other models and with the same dataset. Accuracy and macro F1 score were recorded for each of the 6 fine-tuning sessions and post festum analyzed.
20
+
21
+ | model | average accuracy | average macro F1|
22
+
23
+ |---|---|---|
24
+
25
+ |bcms-bertic-frenk-hate|0.8313|0.8219|
26
+
27
+ |EMBEDDIA/crosloengual-bert |0.8054|0.796|
28
+
29
+ |xlm-roberta-base |0.7175|0.7049|
30
+
31
+
32
+
33
+ From recorded accuracies and macro F1 scores p-values were also calculated:
34
+
35
+ Comparison with`crosloengual-bert`:
36
+
37
+ | test | accuracy p-value | macro F1 p-value|
38
+ | --- | --- | --- |
39
+ |Wilcoxon|0.00781|0.00781|
40
+ |Mann Whithney|0.00108|0.00108|
41
+ |Student t-test |2.43e-10 |1.27e-10|
42
+
43
+ Comparison with`xlm-roberta-base`:
44
+
45
+ | test | accuracy p-value | macro F1 p-value|
46
+ | --- | --- | --- |
47
+ |Wilcoxon|0.00781|0.00781|
48
+ |Mann Whithney|0.00107|0.00108|
49
+ |Student t-test |4.83e-11 | 5.61e-11 |