Enriched BERT Embeddings for Scholarly Publication Classification

TL;DR

This paper explores enriching datasets and fine-tuning pre-trained language models, especially SPECTER2, to improve automatic classification of scholarly articles into research fields, achieving a weighted F1-score of 0.7415.

Contribution

It demonstrates that dataset enrichment and fine-tuning pre-trained models significantly improve scholarly publication classification accuracy.

Findings

Fine-tuning models like SPECTER2 yields the highest accuracy.

Enriching datasets with metadata from bibliographic sources improves classification.

Best model achieves a weighted F1-score of 0.7415.

Abstract

With the rapid expansion of academic literature and the proliferation of preprints, researchers face growing challenges in manually organizing and labeling large volumes of articles. The NSLP 2024 FoRC Shared Task I addresses this challenge organized as a competition. The goal is to develop a classifier capable of predicting one of 123 predefined classes from the Open Research Knowledge Graph (ORKG) taxonomy of research fields for a given article.This paper presents our results. Initially, we enrich the dataset (containing English scholarly articles sourced from ORKG and arXiv), then leverage different pre-trained language Models (PLMs), specifically BERT, and explore their efficacy in transfer learning for this downstream task. Our experiments encompass feature-based and fine-tuned transfer learning approaches using diverse PLMs, optimized for scientific tasks, including SciBERT, SciNCL, and SPECTER2. We conduct hyperparameter tuning and investigate the impact of data augmentation from bibliographic databases such as OpenAlex, Semantic Scholar, and Crossref. Our results demonstrate that fine-tuning pre-trained models substantially enhances classification performance, with SPECTER2 emerging as the most accurate model. Moreover, enriching the dataset with additional metadata improves classification outcomes significantly, especially when integrating information from S2AG, OpenAlex and Crossref. Our best-performing approach achieves a weighted F1-score of 0.7415. Overall, our study contributes to the advancement of reliable automated systems for scholarly publication categorization, offering a potential solution to the laborious manual curation process, thereby facilitating researchers in efficiently locating relevant resources.