sadanyh commited on
Commit
cf85362
1 Parent(s): 8cc476c

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +51 -1
README.md CHANGED
@@ -15,4 +15,54 @@ widget:
15
  example_title: "خليجي"
16
  - text: "و حضرتك طيبة و شكرا علي الكلام الحلو ده يا مبهجة..."
17
  example_title: "مصري"
18
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
15
  example_title: "خليجي"
16
  - text: "و حضرتك طيبة و شكرا علي الكلام الحلو ده يا مبهجة..."
17
  example_title: "مصري"
18
+ ---
19
+ # Dialectical-MSA-detection
20
+
21
+ ## Model description
22
+
23
+ This model was trained on 108,173 manually annotated User-Generated Content (e.g. tweets and online comments) to classify the Arabic language of the text into one of two categories: 'Dialectical', or 'MSA' (i.e. Modern Standard Arabic).
24
+
25
+
26
+ ## Training data
27
+
28
+ Dialectical-MSA-detection was trained on the English-speaking subset of the [The Arabic online commentary dataset (Zaidan, et al 20211)](https://github.com/sjeblee/AOC).
29
+ The AOC dataset was created by crawling the websites of three Arabic newspapers, and extracting online articles and readers' comments.
30
+
31
+
32
+
33
+ ## Training procedure
34
+
35
+ `xlm-roberta-base` was trained using the Hugging Face trainer with the following hyperparameters.
36
+
37
+ ```
38
+ training_args = TrainingArguments(
39
+ num_train_epochs=4, # total number of training epochs
40
+ learning_rate=2e-5, # learning rate
41
+ per_device_train_batch_size=32, # batch size per device during training
42
+ per_device_eval_batch_size=4, # batch size for evaluation
43
+ warmup_steps=0, # number of warmup steps for learning rate scheduler
44
+ weight_decay=0.02, # strength of weight decay
45
+
46
+ )
47
+ ```
48
+
49
+ ## Eval results
50
+
51
+ The model was evaluated using 10% of the sentences (90-10 train-dev split). Accuracy 0.88 on the dev set.
52
+
53
+
54
+ ## Limitations and bias
55
+
56
+ The model was trained on sentences from the online commentary domain. Other forms of UGT such as tweet can be different in the degree of dialectness.
57
+
58
+
59
+ ### BibTeX entry and citation info
60
+
61
+ ```bibtex
62
+ @article{saadany2022semi,
63
+ title={A Semi-supervised Approach for a Better Translation of Sentiment in Dialectical Arabic UGT},
64
+ author={Saadany, Hadeel and Orasan, Constantin and Mohamed, Emad and Tantawy, Ashraf},
65
+ journal={arXiv preprint arXiv:2210.11899},
66
+ year={2022}
67
+ }
68
+ ```