This study proposes a knowledge graph (KG) construction method integrating citation information to address credibility tracing of knowledge claims. By fine-tuning large language models (LLMs), SPO triples are precisely extracted from the "Conclusion" sections of literature abstracts. A dual-level fusion mechanism is developed based on sentiment analysis (Positive/Neutral/Negative) of citation contexts. This approach injects academic evaluation attributes at the paper level and links them to SPO triples. An innovative Neo4j-MySQL heterogeneous storage architecture is designed, enabling fine-grained evidence tracing from KG relationships to citation contexts through uniform identifiers. The constructed KG simultaneously provides knowledge claims and supports credibility tracing from the academic community perspective.
Biomedical literature serves as a critical carrier of scientific research achievements, documenting major
breakthroughs, knowledge discoveries, and direct empirical evidence from clinical studies. Current
studies typically construct domain knowledge graphs by extracting entity relationships from titles and
abstracts[
To overcome these limitations and leverage citation sentiment, this study proposes a KG framework integrating evidence from citation information (including sentiment and context). Core innovations include:
1. Structured Evidence Screening: SPO triples are strictly extracted from "Conclusion" sections of scientific abstracts, ensuring precise reflection of research claims while minimizing information noise.
2. Citation Information Fusion: Deep learning models parse sentiment polarity in original citation contexts to quantify academic evaluations between papers. Citation sentiments, contexts, and SPO triples are fused to enable credibility tracing. These mechanisms aim to transform static KGs into tools supporting knowledge claim credibility tracing.
These mechanisms transform static KGs into tools supporting knowledge claim credibility tracing. The code and data used in our Knowledge Graph all available at https://github.com/wang-lch/CE-KG.
This study establishes a multimodal biomedical knowledge base integrating three heterogeneous data
sources: Metadata of 9,201 clinical research papers on breast cancer drug therapy (including trials
and guidelines) published from 2000-2024 were retrieved from PubMed; 83,510 medical SPO triples
were extracted from the SemMedDB[
An evidence-oriented filtering strategy was adopted, using only Conclusion-derived SPO triples to
ensure claims reflect empirical findings. For confidence quantification, the Dict-Senti-BERT[
A Neo4j-MySQL heterogeneous storage architecture enables dynamic knowledge network storage and ifne-grained evidence tracing. Neo4j stores normalized entities (e.g., drugs, diseases) as nodes and SPO predicates (e.g., TREATS, CAUSES) as directed relationships with embedded evidence support (PMID lists and sentiment distributions). MySQL maintains structured evidence: The Papers table records metadata (titles, structured abstracts), while the Citations table stores citation types and context snippets. A unified identifier system (Triple_ID/PMID/Citation_ID) enables cross-database mapping from KG relationships to specific citation contexts, forming an integrated solution supporting knowledge evolution and claim verification.
This study designs a hierarchical knowledge graph (KG) visualization scheme enabling traceable analysis from knowledge claims to microscopic evidence through three progressive interfaces. The workflow begins at the semantic triple layer where clicking a predicate edge of any SPO triple triggers a detailed attribute panel. This panel displays the PMID list of associated literature. The system then dynamically renders a literature overview layer (central expansion panel). Each paper presents structured data including title, PMID, publication year, and sentiment polarity distribution (e.g., positive/negative citation counts). Finally, the citation layer comprehensively displays all citation contexts of individual papers: Total citation statistics, sentiment classification data, and sentiment-labeled citation contexts with associated metadata (source PMID, title, year) are systematically presented. This forms a complete evidence tracing chain, facilitating credibility verification of knowledge claims.
This study proposes a knowledge graph (KG) construction method integrating citation sentiment and contextual evidence. The evidence-oriented mechanism—strictly extracting SPO triples only from abstract "Conclusion" sections—significantly improves knowledge claim accuracy. Innovatively incorporating citation information enables credibility tracing for academic knowledge discovery. The constructed KG links knowledge claims to source literature’s citation data, allowing examination of academic community evaluations and facilitating access to high-credibility claims.
Future work will employ deep learning tools to extract multi-source citation data, complementing the Colil database to enrich academic evaluations. An alignment mechanism between citation contexts and SPO triples will be explored for precise evidence tracing. Additionally, applicability assessment in other domains will be conducted.
This research was funded by the Innovation Fund for Medical Sciences of the Chinese Academy of Medical Sciences (Grant number: 2021-I2M-1-033)
During the preparation of this manuscript, the authors employed ChatGPT and Grammarly to support grammar correction, spelling refinement, and overall language polishing. These tools were used to enhance clarity and professionalism in writing. The authors subsequently reviewed and revised all content, and took full responsibility for the accuracy and integrity of the final publication.