Nevise: A Bert-Based Spell-Checker for Persian

Nevise is a Persian spelling-checker developed by Dadmatech Co based on deep learning. Nevise is available in two versions. The second version has greater accuracy, the ability to correct errors based on spaces, and a better understanding of special characters like half space. These versions can be accessed via web services and as demos. We provide public access to the code and model checkpoint of the first version here.

packages Installation

Use the package manager pip to install packages.

pip install -r requirements.txt

Download model checkpoint and vocab and put them on "model" directory

pip install gdown
mkdir model
cd model
gdown https://drive.google.com/uc?id=1Ki5WGR4yxftDEjROQLf9Br8KHef95k1F
gdown https://drive.google.com/uc?id=1nKeMdDnxIJpOv-OeFj00UnhoChuaY5Ns

run

python main.py

Demo

Nevise(both versions)

Results on Nevise Dataset

Algorithm	Wrong detection rate	Wrong correction rate	Correct to wrong rate	Precision
Nevise 2	0.8314	0.7216	0.003	0.968
Paknevis	0.7843	0.6706	0.228	0.7921
Nevise 1	0.7647	0.6824	0.0019	0.9774
Google	0.7392	0.702	0.0045	0.9449
Virastman	0.6	0.5	0.0032	0.9533