Enhancing topic modeling in digital libraries through keyword-centric self-supervised learning

Abstract

In the present ever-changing information environment, libraries are encountering both opportunities and challenges. The adoption of information technologies has transformed how libraries identify, acquire, process, and distribute content to users. However, the integration of digital systems also creates immediate pressures for both librarians and library users. The academic library environment undergoes a shift, necessitating a focus on key trends that align with its fundamental characteristics. This paper provides a comprehensive analysis of digital libraries from a research perspective, exploring essential characteristics, emerging trends, and areas in need of further investigation. We introduce a novel end-to-end keyword-centric topic modeling approach utilizing a Self-supervised learning graph convolutional network (GCN) for topic discovery. This approach analyzes a collection of publications from SCOPUS in the digital libraries domain. A self-supervised learning technique is applied via a link prediction task to learn the semantic representation of keywords, enabling highly accurate keyword clustering for topic modeling without the need for manually labeled data. To evaluate the proposed framework, we established two baseline models: FastText embedding with k-means and Girvan-Newman clustering, alongside human annotation. The proposed keyword-centric model achieved an accuracy of 0.84 and an ARI of 0.19, surpassing the baseline models. The exploration of eight topics revealed significant features of studies in the digital libraries field, highlighting areas predominantly explored and those requiring more research attention. The essential contribution of this research lies in its keyword-based investigation, which mitigates inaccuracies inherent in traditional text-based topic modeling. By integrating advanced methodologies, this study bridges the gap between cutting-edge technological capabilities and practical applications in digital libraries, demonstrating how contemporary artificial intelligence can enhance topic modeling.