![Screenshot 2023-12-18 at 15 50 10](https://private-user-images.githubusercontent.com/16616024/291303324-5d1c9b9b-59bc-4efe-a66b-83fbd2650842.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.ABhdrjB5Y8AR3TY73XCNlBiy9mqRoT1DIlcIqbj0MOw)
AdapterFusion is a learning algorithm that combines task-relevant information contained within adapters to fine-tune a target task using multiple resources. We employ AdapterFusion on a classsic Arabic NLP problem, Dialect Identification, and deploy the resulting model on HuggingFace Spaces🤗
This project is a fraction of a Msc degree dedicated to Arabic Dialect Identification that was supervised by Prof. Dr. Mervat Abouelkhair
- An AdapterFusion model requires N adapters
- Each n adapter is fine-tuned on task-relevant data
- In our case, we fine-tune, we fine-tune a total of 8 adapters on the data below covering MSA/DA, Region, Country and Province Levels
- For each diatopic variation an AdapterFusion (AF) model is created, hence AF for each Region, Country, Province and MSA/DA
- We share one of the best performing AF models on Regional Dialect Identification here: HuggingFace Spaces🤗
- Arabic Online Commentary (AOC), Newspaper Commentary
- NADI 2020, Tweets
- MADAR, Tweets
- ArSarcasm, Tweets
- QADI, Tweets
Code for adapter-training and adapterfusion available in repo for re-producing