NeuralDialectDetector

We achieve state of the art results on all of the 2021 shared NADI tasks by ensembling a set of models built on top MARBERT (https://arxiv.org/abs/2101.01785) with different architectural variants.

Table of Results

Subtask	Models	DEV	DEV	TEST	TEST
		Acc.	F1	Acc.	F1
1.1	Ours	39.06	23.52	35.72	22.38
1.2	MARBERT	48.86	26.40	48.40	29.14
1.2	Ours	53.42	34.03	51.66	32.26
2.1	Ours	7.04	6.73	6.66	6.43
2.2	MARBERT	7.91	5.23	8.48	6.28
2.2	Ours	10.74	10.02	9.46	8.60

Link to Paper

Website Demo

A demo of the model's predictions is available on a WebApp: https://arabicdialectdetector.azurewebsites.net