# The Effect of Metadata on Scientific Literature Tagging: A Cross-Field Cross-Model Study

**Authors:** Yu Zhang, Bowen Jin, Qi Zhu, Yu Meng, Jiawei Han  
**arXiv:** [2302.03341](https://arxiv.org/abs/2302.03341)  
**Date:** 2023-02-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2302.03341v1/

## Key points

- The necessity to categorize scientific papers with precise topics has become increasingly urgent in the rapidly expanding landscape of scientific publications on the Web.
- Scientific literature tagging process goes beyond simple multi-label text classification and includes utilizing metadata such as details about venues, authors, and references to determine relevant tags for each paper.
- Previous studies have been limited in scope and focused on specific scientific fields like computer science and biomedicine, often using only one specific model.
- This study takes a comprehensive approach by investigating the impact of metadata on scientific literature tagging across 19 different fields using three representative multi-label classifiers: a bag-of-words model, a sequence-based model, and a pre-trained language model.
- Analysis revealed common patterns in the effects of metadata across all fields, with venues consistently proving beneficial for paper tagging in nearly all cases.
- Unique patterns were uncovered in fields outside of computer science and biomedicine that had not been previously explored.
- Metadata plays a significant role in enhancing the accuracy and efficiency of scientific literature tagging across diverse academic disciplines.

## Summary

In the rapidly expanding landscape of scientific publications on the Web, the necessity to categorize each paper with precise topics has become increasingly urgent. This enables researchers to focus on their specific areas of interest without being overwhelmed by the sheer volume of literature available. The process of scientific literature tagging goes beyond simple multi-label text classification, as papers online are often accompanied by a wealth of metadata including details about venues, authors, and references. These additional pieces of information can serve as valuable signals in determining relevant tags for each paper. While previous studies have explored the use of metadata in academic paper classification, they have typically been limited in scope and focused on just one or two scientific fields such as computer science and biomedicine. They also often only use one specific model. In contrast, this study takes a comprehensive approach by systematically investigating the impact of metadata on scientific literature tagging across 19 different fields. The research team selected three representative multi-label classifiers - a bag-of-words model, a sequence-based model, and a pre-trained language model - to assess how their performance in scientific literature tagging is influenced when metadata is incorporated as additional features. Through their analysis, they identified some common patterns in the effects of metadata across all fields; for example, venues consistently prove beneficial for paper tagging in nearly all cases. Additionally, they uncovered unique patterns in fields outside of computer science and biomedicine that had not been previously explored. Overall, this study sheds light on the significant role that metadata plays in enhancing the accuracy and efficiency of scientific literature tagging across diverse academic disciplines. By leveraging these additional signals effectively, researchers can more effectively navigate the vast sea of scholarly publications available online.