COVID Twitter Sentiment Classifier

This repo contains our fine-tuned BERT sentiment classifier for COVID tweets. Our model adopts and finetunes the Covid-Twitter Bert model (Müller et al., 2020).

You can easily use our off-the-shelf model by a simple 1-line command.

Main author of this repo: Tejas Vaidhya. This is part of the project on NLP for Policy-Making led by Zhijing Jin (Max Planck Institute), Zeyu Peng (MIT), Tejas Vaidhya (IIT), Bernhard Schoelkopf (Max Planck Institute), and Rada Mihalcea (University of Michigan).

In this repo, we provide the following resources:

• Our finetuned BERT model (and model weights saved here)
• Command to run our BERT model (with both interactive and inference modes)
• Dependencies and step-by-step guides

Directory Structure

• data: Contains pre-processed csv file of data. Read data/README.md for more details.

• experiment: Contains experiments related log and models

• pretrained_model: Pretrained models and config jsons

• env.yml: yaml file to create conda environment

• inference.py: code for runing model in inference mode

• interactive.py: code to run model in interactive mode

• README: This file :)

• train.py: code for training the Covid twitter bert

• util.py: contains utilty function

Instructions

1. Set up the codebase and the dependencies

# creating conda environment
conda env create --file env.yml
# clone this repo
git clone https://github.com/tejasvaidhyadev/covid-twitter-sentiment-classification.git

Below are the dependencies of our model:

2. Download the pretrained model

Download pretrained model and keep it in pretrained_model folder

wget https://github.com/tejasvaidhyadev/covid-twitter-sentiment-classification/releases/download/V0.1/weights.zip
unzip weights.zip -d pretrained_model #make sure to del already existed JSON.  

3. Run Our Model by a Simple One-Line Command

python interactive.py --nolog True --trained_model ${path_to_trained_model} --pretrained_dir ${path_to_pretrained_model}

Example to run our model in the interactive mode:

$ python 

interactive.py --nolog True
loading pretraining weights
======= ======= welcome! ======== ======= 
    1. Provide input strings
    2. Input 'exit' to end

Input:Input:NEW REPORT: 85% OF PEOPLE WHO CAUGHT COVID WERE REGULAR MASK WEARERS!! So masks dont do a thing to stop the virus!! TOTAL BS!! Dems using it to CONTROL US!! #TakeOffYourMasksSheeple UNHEALTHY FOR HEALTHY PEOPLE TO WEAR MASKS!!!
NEWSOM LOOKS LIKE A TOTAL MORON
['negative']
Input:  

4. Inference mode

Read the data/README.md and create the csv file of desired formate

python inference.py -csv ${path_to_csv}

Miscellanous

• This project is MIT-Licenced.
• To request new features, please start a pull request.