COVHINDIA : Deep Learning Model for Sentiment Classification on COVID19 Tweets from India

Overview

This repository describes a framework to perform sentiment analyis on COVID-19 tweets posted in Hindi language on Twitter platform. The framework leverages open-source machine translation tools to translate Hindi tweet to English and then pass the preprocessed translated tweet as an input to a BERT-based model for performing multi-lingual sentiment polarity detection.

Datasets

1. Tweet dataset : Kaggle dataset
2. GloVe embeddings : GloVe
3. FastText embeddings : FastText
4. GloVe Twitter : GloveTwitter
5. Crisis Embeddings : Crisis Embeddings

Model Description	Training Accuracy	Validation Accuracy	Notebook Link
Basic LSTM	85.1%	86.7%	Notebook
LSTM + GloVe Embeddings	88.9%	90.9%	Notebook
LSTM + FastText Embeddings	92.5%	88.9%	Notebook
LSTM + Crisis Embeddings	83.4%	84.7%	Notebook
Basic Bi-directional LSTM	87.3%	86.0%	Notebook
Bi-directional LSTM + GloVe Embeddings	91.2%	90.6%	Notebook
Bi-directional LSTM + FastText Embeddings	88.3%	88.6%	Notebook
Bi-directional LSTM + Crisis Embeddings	86.0%	85.1%	Notebook
BERT	99.7%	93.8%	Notebook

System Architecture

Below is the system architecture for sentiment polarity detection of COVID-19 tweets in Hindi using machine translation and BERT.

Research Paper

I have published my findings as a research paper: 'Covhindia: Deep Learning Framework for Sentiment Polarity Detection of Covid-19 Tweets in Hindi' in the 'International Journal on Natural Language Computing'

References

Paper on sentiment analysis

Using NMT/HSWN for sentiment analysis on Hindi tweets