Heterogeneous deep graph convolutional network with citation relational BERT for COVID-19 inline citation...

Area of Science:

• Computational linguistics
• Bibliometrics
• Artificial Intelligence in Scientific Research

Background:

• The COVID-19 pandemic has led to an unprecedented surge in scientific publications.
• Navigating this vast amount of literature to find relevant citations is challenging, particularly for early-career researchers.
• Efficient citation recommendation systems are crucial for staying current in rapidly evolving research fields.

Purpose of the Study:

• To develop a novel method for improving inline citation recommendation in COVID-19 research.
• To address the challenge of information overload in the scientific literature generated by the pandemic.
• To enhance the ability of researchers to discover relevant cited papers within documents.

Main Methods:

• Introduction of a new neural network model: citation relational BERT with heterogeneous deep graph convolutional network (CRB-HDGCN).
• Stage 1: Enhancing BERT model representation learning using an augmented citation sentence corpus and fine-tuning with citation relation prediction tasks.
• Stage 2: Employing a heterogeneous deep graph convolutional network (HDGCN) with multiple layers (embedding, heterogeneous graph convolution, attention, decoder) to learn dense node vectors from a COVID-19 bibliographic graph.

Main Results:

• The CRB-HDGCN model significantly improved citation recommendation performance on both CORD-19 and LitCovid datasets.
• Compared to the CO-Search method, CRB-HDGCN demonstrated substantial percentage increases in Mean Average Precision (MAP), Mean Reciprocal Rank (MRR), Precision at 100 (P@100), and Recall at 100 (R@100).
• Specific improvements on CORD-19 dataset: MAP (21.8%), MRR (22.7%), P@100 (37.6%), R@100 (21.2%).
• Specific improvements on LitCovid dataset: MAP (29.1%), MRR (25.9%), P@100 (15.3%), R@100 (11.3%).

Conclusions:

• The proposed CRB-HDGCN method offers a powerful solution for COVID-19 inline citation recommendation.
• The model effectively leverages BERT and heterogeneous graph convolutional networks to enhance literature discovery.
• This approach can significantly aid researchers in efficiently identifying relevant scientific literature amidst information overload.