From Literature Overload to Knowledge Graph: An Automated Pipeline for Literature Reviews

Abstract

The rapid expansion of scientific publications has significantly increased the complexity of traditional literature review processes. While recent advances in AI-assisted screening reduce manual effort, they fail to provide an actionable organization of findings beyond thematic clustering. We propose an integrated pipeline that transforms raw bibliographic data into queryable Knowledge Graphs (KGs), combining: (1) automated collection via the OpenAlex API, (2) LLM-assisted screening, (3) hierarchical semantic clustering using state-of-the-art embeddings (Qwen3-Embedding-4B), and (4) multi-relational KG construction in Neo4j with GraphRAG. We validate this methodology on a corpus of 50 K articles on Artificial Intelligence from the computer science literature. Our hierarchical clustering identifies 7 macro-clusters and 117 microclusters with Fused Gromov-Wasserstein (FGW) coherence. The resulting KG integrates 34,200 nodes across 9 entity types, revealing temporal evolution patterns, cross-institutional collaborations, and foundational knowledge pillars through citation analysis.