Large Scale Subject Category Classification of Scholarly Papers With Deep Attentive Neural Networks

Area of Science:

• Bibliometrics
• Information Science
• Computer Science

Background:

• Scholarly paper subject category classification is crucial for bibliometrics and digital libraries.
• Existing methods often rely on citation networks, which are unavailable for new or uncited papers.
• A significant gap exists in classifying papers lacking metadata or citation information.

Purpose of the Study:

• To develop a deep attentive neural network (DANN) for classifying scholarly papers using only their abstracts.
• To address the limitations of citation-based classification methods, particularly for papers with sparse citation data.
• To evaluate the performance of the DANN model across 104 subject categories.

Main Methods:

• A deep attentive neural network (DANN) architecture comprising two bi-directional recurrent neural networks and an attention layer was proposed.
• The model was trained on nine million abstracts from Web of Science (WoS), utilizing the WoS schema of 104 subject categories.
• Comparative analysis was performed against baseline models, varying network architecture and text representation (word vectors with TFIDF, character, and sentence embeddings).

Main Results:

• The DANN model achieved a micro-averaged F1-score of 0.76, with category-specific F1-scores ranging from 0.50 to 0.95.
• Retraining word embedding models significantly improved vocabulary overlap and model performance.
• The attention mechanism proved effective, and the combination of word vectors with TFIDF outperformed other embedding strategies.

Conclusions:

• The proposed DANN model offers a robust solution for subject category classification of scholarly papers based solely on abstracts.
• The study highlights the importance of tailored word embeddings and attention mechanisms for enhanced classification accuracy.
• The findings provide a foundation for improving scientific literature organization and discoverability, especially for citation-scarce publications.