COVID Twitter Misinformation Lookup: An Investigation of Relevance Ranking Methods

Team members: Swati Adhikari, Provakar Mondal, Kiet Nguyen

This repository contains the code for our CS 5604 (Information Storage and Retrieval) final project at Virginia Tech.

Semester: Fall 2021

Instructor: Dr. Ismini Lourentzou

In this course project, we aim to build a search engine using COVID-19-related Tweets as the input data; we experiment with 3 different models: Okapi BM25, RankNet, and TF-Ranking.

Dataset

The dataset we use for this project is borrowed from this paper:

@misc{dharawat2020drink,
      title={Drink bleach or do what now? Covid-HeRA: A dataset for risk-informed health decision making in the presence of COVID19 misinformation},
      author={Arkin Dharawat and Ismini Lourentzou and Alex Morales and ChengXiang Zhai},
      year={2020},
      eprint={2010.08743},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Click here for more detailed instructions on downloading the data.

Requirements

The requirements were generated with pip freeze > requirements.txt and can be found in requirements.txt.

Getting data and training models

1. Create a new virtual environment and install the packages noted in requirements.txt.
2. Set up your Twitter API keys (as detailed here).
3. Run get_data.sh to obtain raw tweets with the Twitter API. This step could take a long time (up to days) due to the Twitter API rate limit of 900 requests per 15 minutes. However, you can always continue where you left off (i.e. this code would not overwrite downloaded data if you stopped it and ran it again).
4. Run prepare_data.sh to prepare input data for the 3 different models.
5. Run any of the three files in models with python {filename} or python3 {filename} to run or train the models, or run run_models.sh to execute all 3 models sequentially.