Efficient Self-Supervised Metric Information Retrieval: A Bibliography Based Method Applied to COVID Literature

Area of Science:

• Information Retrieval
• Computational Linguistics
• Bibliometrics

Background:

• The vast and growing body of scientific literature, particularly on coronaviruses (over 300,000 publications), necessitates efficient methods for knowledge discovery.
• Current information retrieval (IR) systems often rely on deep learning with supervised training, requiring extensive labeled datasets and expert input, which are resource-intensive and slow to create, especially during urgent research periods like a pandemic.
• The COVID-19 Open Research Dataset (CORD19) presents a significant corpus for evaluating IR systems.

Purpose of the Study:

• To develop a novel, self-supervised information retrieval (IR) engine capable of searching scientific paper corpora for relevant documents against arbitrary queries without requiring pre-labeled datasets.
• To evaluate the performance of this new system, named SUBLIMER, against state-of-the-art competitors on the CORD19 dataset.

Main Methods:

• Developed SUBLIMER, a self-supervised IR engine utilizing deep metric learning.
• Trained SUBLIMER on the unsupervised CORD19 dataset.
• Exploited the citation network of papers to create a latent space where proximity indicates semantic similarity, eliminating the need for labeled query-document pairs.

Main Results:

• SUBLIMER demonstrates superior performance compared to state-of-the-art competitors on the CORD19 dataset, specifically in Precision@5 (P@5) and Bpref metrics.
• The self-supervised approach of SUBLIMER significantly outperforms supervised methods that require labeled data.
• SUBLIMER achieves this high performance with an order of magnitude fewer trainable parameters than competing approaches.

Conclusions:

• SUBLIMER offers an efficient and effective self-supervised solution for information retrieval in large scientific literature corpora.
• The methodology, leveraging citation networks for semantic similarity, is adaptable to other scientific domains beyond CORD19.
• This approach addresses the limitations of traditional supervised IR methods, particularly in resource-constrained or time-sensitive research scenarios.