NLP for Bengali
August 6, 2020 ยท View on GitHub
This repository contains State of the Art Language models and Classifier for Bengali language, which is primarily spoken by the Bengalis in South Asia.
The models trained here have been used in Natural Language Toolkit for Indic Languages (iNLTK)
Dataset
Created as part of this project
Open Source Datasets
Results
Language Model Perplexity (on validation set)
| Architecture/Dataset | Bengali Wikipedia Articles |
|---|---|
| ULMFiT | 41.2 |
| TransformerXL | 39.3 |
Classification Metrics
ULMFiT
| Dataset | Accuracy | MCC | Notebook to Reproduce results |
|---|---|---|---|
| Bengali News Articles (Soham Articles) | 90.71 | 87.92 | Link |
Visualizations
Word Embeddings
| Architecture | Visualization |
|---|---|
| ULMFiT | Embeddings projection |
| TransformerXL | Embeddings projection |
Results of using Transfer Learning + Data Augmentation from iNLTK
On using complete training set (with Transfer learning)
| Dataset | Dataset size (train, valid, test) | Accuracy | MCC | Notebook to Reproduce results |
|---|---|---|---|---|
| Bengali News Articles (Soham Articles) | (11284, 1411, 1411) | 90.71 | 87.92 | Link |
On using 1% of training set (with Transfer learning)
| Dataset | Dataset size (train, valid, test) | Accuracy | MCC | Notebook to Reproduce results |
|---|---|---|---|---|
| Bengali News Articles (Soham Articles) | (112, 1411, 1411) | 69.88 | 61.56 | Link |
On using 1% of training set (with Transfer learning + Data Augmentation)
| Dataset | Dataset size (train, valid, test) | Accuracy | MCC | Notebook to Reproduce results |
|---|---|---|---|---|
| Bengali News Articles (Soham Articles) | (112, 1411, 1411) | 74.06 | 65.08 | Link |
Pretrained Models
Language Model
Download pretrained Language Model from here
Tokenizer
Trained tokenizer using Google's sentencepiece
Download the trained model and vocabulary from here