An ensemble approach for research article classification: a case study in artificial intelligence.

Area of Science:

• Computer Science
• Information Science

Background:

• Text classification in emerging scientific fields is challenging due to evolving boundaries and interdisciplinary nature.
• Traditional keyword-based methods suffer from low recall due to incomplete term lists.

Purpose of the Study:

• To develop and evaluate a deep learning-based ensemble approach for accurate article classification in dynamic research areas.
• To address the limitations of traditional methods in capturing the full scope of emerging scientific fields, using artificial intelligence (AI) as a case study.

Main Methods:

• An ensemble model combining decision tree, SciBERT, and regular expression matching was developed.
• A support vector machine (SVM) was used to integrate results from individual models.
• The approach was evaluated on a manually labeled dataset from the Web of Science (WoS) corpus.

Main Results:

• The ensemble model achieved 97% recall and 0.92 precision in identifying AI-related articles.
• This represents a 0.15 increase in F1-score compared to existing search term-based approaches.
• An ablation study confirmed the contribution of each ensemble component, with SciBERT showing superior performance over other BERT models.

Conclusions:

• The proposed deep learning ensemble method effectively enhances the classification of research articles in rapidly evolving scientific domains.
• This approach offers a significant improvement over traditional methods, particularly for interdisciplinary and emerging fields.
• SciBERT demonstrates strong efficacy within the ensemble for identifying relevant scientific literature.