CausalCite: A Causal Formulation of Paper Citations

This repo contains the code for the paper:

CausalCite: A Causal Formulation of Paper Citations (2023)

Ishan Kumar*, Zhijing Jin*, Ehsan Mokhtarian, Siyuan Guo, Yuen Chen, Negar Kiyavash, Mrinmaya Sachan, Bernhard Schoelkopf (*: Co-first author)

Usage

1. Download the Semantic Scholar Dataset

To download the Dataset you need a Semantic scholar API key which lets you make 5k requests in 5 mins. This key can be obtained by filling out the form on the Semantic Scholar Website. https://www.semanticscholar.org/product/api#api-key-form. Once you have the key, replace it in the downloadDatasets.py file in the apiKey value.

datasetDownload = "papers"
python downloadDatasets.py

datasetDownload = "abstracts"
python downloadDatasets.py

The Datasets will be downloaded in the form of 30 .jsonl files in the mentioned output folder

2. Preprocess the Data

The 3 datasets Abstracts, Papers, Citations can be preprocessed using the scripts in code/dataProc/*.py. This will combine the relevant data from the 30 jsonl files into 1 single .csvfile.

3. Run the CausalCite Code

Once all the Datasets are created, you can run the causal cite code by replacing the path variables.

python code/getCIIdata.py --model_path /path/to/specter2 --citationData_path /path/to/fullCitationDataset.parquet.gzip --paperData_path /path/to/fullpapersdata.parquet.gzip --bm25Path /path/to/bm25_data --output_path /path/to/output --CandidatePool_path /path/to/candidate_pool --paperAid 3303339

The output will be created as a folder named as PaperA's ID and it will contain all of its sampled Citation Paper Bs files. These can be used to calculate the TCI.