What makes a knowledge graph useful for citation checking?
A knowledge graph helps citation-checking AI by giving it more than just a list of 'paper A cites paper B' — it adds context like the purpose of the citation and hidden relationships between papers. The 2022 GLAD model explicitly incorporated both node attributes (paper metadata) and link attributes (why a citation was made) via graph neural networks, and it detected anomalous citations on a simulated dataset [1]. The model's 'Citation Purpose' (CPU) algorithm mined the text around each citation to classify its intent, which turned out to be critical for spotting manipulation [1]. Without that purpose information, the graph is just a network of links, and the AI has a much harder time distinguishing a legitimate citation from a gamed one.
How does graph structure help flag manipulated citations?
The structure of the citation graph itself — not just the individual links — can reveal manipulation. A 2023 study introduced a method that deliberately perturbed (slightly altered) the citation graph and then used a deep embedding model to see which connections held up under stress [2]. The idea is that legitimate citations form robust, redundant pathways in the graph, while manipulated ones are fragile and disappear when the network is perturbed. In numerical experiments, this approach reliably identified a core set of trustworthy citations and flagged suspicious references that were likely manipulated [2]. This works because the graph captures indirect relationships — a manipulated citation often lacks the supporting web of indirect connections that a real citation would have.
Where do knowledge graphs fall short, and what else is needed?
A knowledge graph alone isn't enough — the AI still needs high-quality entity extraction and deeper relationship modeling. A 2025 project built a legal knowledge graph from 8,400 Congolese legal documents using LLM-based named entity recognition, and it successfully created a preliminary citation graph with over 1,400 relationships [3]. However, the authors explicitly noted that the current methodology only handled basic entity extraction (document types, reference numbers, dates) and that future work needed deeper relationship modeling and better type coverage to make the graph truly useful for automated citation checking [3]. This shows that building the graph is just the first step; the AI's ability to check citations depends on how well the graph captures the semantics of each citation.
About These Sources
This answer is built on 4 peer-reviewed studies — published from 2021 to 2025, 1 from 2024 or later, 3 in Q1 journals, collectively cited 80 times — selected as the most relevant from 4 studies that passed quality screening, drawn from 54 papers retrieved from a database of over 500 million.
Sources used in this answer
Deep Graph Learning for Anomalous Citation Detection
Proposed GLAD model that combines text semantics with graph neural networks to detect anomalous citations; includes a 'Citation Purpose' algorithm that mines citation context to classify why a citation was made, and outperformed baselines on a simulated anomalous citation dataset.
Detecting Pseudo-Manipulated Citations in Scientific Literature through Perturbations of the Citation Graph
Introduced a method that perturbs the citation graph and uses deep embeddings to identify reliable citations; numerical experiments showed it could reliably flag manipulated references by exploiting the graph's indirect pathways.
Automated Citation Detection in Congolese Legal Texts: Leveraging LLM-Based NER for Knowledge Graph Construction
Built a legal knowledge graph from 8,400 Congolese legal documents using LLM-based NER, creating a preliminary citation graph of over 1,400 relationships; noted that deeper relationship modeling and improved entity coverage are still needed for effective citation detection.
Data citation and the citation graph
Argued that the citation graph must model 'citable units' and 'reference subsumption' to properly represent data citation and credit flow over time; without these extensions, the graph cannot support accurate bibliometric computations or detect citation gaming.