SciDocFind

Information Retrieval Course Project

Abstract

We propose an end-to-end approach for faceted scientific paper retrieval that employs the SciRepEval dataset and pre-trained language models fine-tuned with contrastive loss using triples generated from the Highly Influential Citations dataset. Experimental results demonstrate competitive performance across different facets of the CSFCube dataset and competitive performance in other facets, addressing the limitations of previous methods.

Project Report

Please check report.pdf

Instructions for execution

• baselines_inference.ipynb Run this notebook* to get the evaluation metrics like R-Precsion, Precision@20, and others for all the baseline models.
  A total of 7 baselines are implemented which consists of 3 abstract level baselines and 4 sentence level baselines.

• finetune_inference.ipynb Run this notebook* after uploading CSFCube-master.zip* to get the evaluation metrics like R-Precsion, Precision@20, and others for all the fine-tuned models. Three models are finetuned - SentBERT-PP, SPECTER and SciNCL.

• sentbert_pp_finetune.ipynb Run this notebook** to fine-tune SentBERT-PP model using PARADE dataset.

• specter_scincl_finetune.ipynb Run this notebook after to fine-tune SPECTER and SciNCL models using the CS papers from HIC dataset.

• demo.ipynb Run this notebook* after uploading CSFCube-master.zip* for an interactive demo of our retrieval system.

• SciRepEval_SciBERT_classifier.ipynb Run this notebook to fine-tune SciBERT model for the task of classifying the field of study of any research paper.

• * upload CSFCube dataset before running the notebook.

• ** upload PARADE_dataset before running the notebook.