mehedihasanbijoy / DPCSpell

A Transformer-based Detector-Purificator-Corrector Framework for Spelling Error Correction of Bangla and Resource Scarce Indic Languages

DPCSpell

A Transformer-based Detector-Purificator-Corrector Framework for Spelling Error Correction of Bangla and Resource Scarce Indic Languages
Preprint — https://arxiv.org/abs/2211.03730

How DPCSpell works?

Running Test

Operating System	Requirement	Remark
Ubuntu 16.04.7 LTS	requirements_u.yml	✔️ Successful
Ubuntu 18.04.6 LTS (Google Colab)	requirements_c.txt	✔️ Successful*
Windows 10	requirements_w.yml	✔️ Successful

Get Started

git clone https://github.com/mehedihasanbijoy/DPCSpell.git

or manually download and extract the github repository of DPCSpell.

Environment Setup

Create A Virtual Environment

conda env create -f requirements_u.yml (for Ubuntu 16.04.7 LTS)
or
conda env create -f requirements_w.yml (for Windows 10)

Activate the Environment

conda activate DPCSpell

Prepare SEC Corpora

gdown https://drive.google.com/drive/folders/1_sWSi-LFsvuYh9c5GBMDd4V6_uM8yYjH?usp=share_link -O ./Dataset --folder

or manually download the folder from here and keep the extracted files into ./Dataset/

Training and Evaluation of DPCSpell

Detector Network

python detector.py --CORPUS "./Dataset/corpus.csv" --HID_DIM 128 --ENC_LAYERS 5 --DEC_LAYERS 5 --ENC_HEADS 8 --DEC_HEADS 8 --ENC_PF_DIM 256 --DEC_PF_DIM 256 --ENC_DROPOUT 0.1 --DEC_DROPOUT 0.1 --CLIP 1 --LEARNING_RATE 0.0005 --N_EPOCHS 100

Purificator Network

python purificator.py --HID_DIM 128 --ENC_LAYERS 5 --DEC_LAYERS 5 --ENC_HEADS 8 --DEC_HEADS 8 --ENC_PF_DIM 256 --DEC_PF_DIM 256 --ENC_DROPOUT 0.1 --DEC_DROPOUT 0.1 --CLIP 1 --LEARNING_RATE 0.0005 --N_EPOCHS 100 

Corrector Network

python corrector.py --HID_DIM 128 --ENC_LAYERS 5 --DEC_LAYERS 5 --ENC_HEADS 8 --DEC_HEADS 8 --ENC_PF_DIM 256 --DEC_PF_DIM 256 --ENC_DROPOUT 0.1 --DEC_DROPOUT 0.1 --CLIP 1 --LEARNING_RATE 0.0005 --N_EPOCHS 100 

Benchmarking Bangla SEC Task