| --- |
| language: |
| - ckb |
| tags: |
| - kurdish |
| - sentiment-analysis |
| - offensive-language-detection |
| - multi-task-learning |
| - kubert |
| --- |
| |
| # KurdSense — Multi-Task Kurdish Sentiment & Offensive Language Detector |
|
|
| KurdSense is the first multi-task NLP model for Kurdish Sorani, simultaneously predicting **sentiment** (Negative/Neutral/Positive) and **offensiveness** (Safe/Offensive) from a single Kurdish tweet. |
|
|
| ## Model Details |
| - **Base model:** KuBERT (asosoft/KuBERT-Central-Kurdish-BERT-Model) |
| - **Architecture:** Shared KuBERT encoder + two classification heads |
| - **Language:** Kurdish Sorani (ckb) |
| - **Dataset:** STSD — 24,668 annotated Kurdish tweets (Wady, Badawi & Kurt, 2024) |
|
|
| ## Results |
|
|
| | Task | Macro F1 | Accuracy | |
| |---|---|---| |
| | Sentiment (3-class) | 0.5100 | 55.9% | |
| | Offensiveness (binary) | 0.6326 | 87.1% | |
|
|
| ## How to Use |
|
|
| Load the model weights using the custom KurdSenseMultiTask architecture |
| and BertTokenizer with unk_token="[UNK]". |
| |
| ## Citation |
| Wady, S. H., Badawi, S., & Kurt, F. (2024). A Kurdish Sorani Twitter dataset |
| for language modelling. Data in Brief, 57, 110967. |
| |
