CorpusLD: A Dual-Layer Semantic Extraction Framework and Deep Knowledge Graph Architecture for Scientific Literature with Deterministic Unit Ontology and Dynamic Authority Disambiguation
Abstract
Scientific and technical documents published in Portable Document Format (PDF) frequently function as unstructured "data graveyards," creating severe semantic isolation. Conventional Retrieval-Augmented Generation (RAG) and large language model (LLM) chunking pipelines suffer from three fundamental failure modes: top-K context truncation loss, multi-page tabular fragmentation, and numeric collisions caused by superscript citation markers polluting quantitative parameters (e.g., misinterpreting reference indices as numerical quantities).
To resolve these challenges, we introduce CorpusLD (Corpus + Linked Data), a production-grade, dual-layer semantic extraction engine and deep knowledge graph architecture. CorpusLD combines a macro-level W3C Schema.org ScholarlyArticle JSON-LD serialization with a micro-level Deep Knowledge Graph ($\mathcal{G} = (\mathcal{V}, \mathcal{E})$) spanning 10 formal semantic predicates. The system incorporates a 4-tier layout-aware parser, a section-wise map-reduce orchestrator, and a deterministic scientific unit ontology supporting Base SI, clinical, energy, and compound units. Furthermore, CorpusLD integrates a dynamic REST authority linker communicating with the Research Organization Registry (ROR v2) and Wikidata/MeSH registries.
Evaluated across an 8-document multi-disciplinary benchmark corpus (arXiv, IEEE, SINTA, and Springer), CorpusLD achieves 100% structural compliance on the Google Rich Results Test, extracts 100% of complex multi-page tables without boundary corruption, and eliminates superscript citation pollution deterministically. CorpusLD is released as an open-source framework and Python package.
Technical Architecture & Validation
CorpusLD processes unstructured multi-disciplinary PDFs through a deterministic 5-stage pipeline:
Sample Cypher query executed on CorpusLD property graph to uncover conflicting methodologies and calibrated metrics across literature without vector re-ranking:
MATCH (p1:Paper)-[:USES_METHOD]->(m1:Method)
MATCH (p2:Paper)-[:USES_METHOD]->(m2:Method)
MATCH (m1)-[r:CONTRADICTS]->(m2)
MATCH (p1)-[:EVALUATES]->(k:Metric)
WHERE k.unit = 'ms' AND k.value < 50.0
RETURN p1.title AS PaperA, m1.name AS MethodA,
r.reason AS ConflictBasis, m2.name AS MethodB,
k.name AS MetricName, k.value AS LatencyVal;
Empirical comparative performance across 8 multi-disciplinary ground-truth benchmark documents:
| Evaluation Metric | Naive Vector RAG Baseline | CorpusLD (Proposed) |
|---|---|---|
| Section Extraction Recall | 54.2% | 100.0% |
| Table Structure Preservation | 37.6% | 100.0% |
| Quantitative Metric Recall | 51.3% | 98.5% |
| Superscript Citation Error Rate | 43.1% | 0.0% (Zero Collision) |
| Context Truncation Loss Rate | 48.7% | 0.0% (Zero Loss) |
| Google Rich Results Structural Pass | 0.0% | 100.0% |
Beyond static document ingestion, CorpusLD establishes the architectural foundation for Living Knowledge Graphs. By decoupling document visualization from binary document storage, CorpusLD federates with open scholarly registries (OpenAlex Works API, Crossref REST API, and Semantic Scholar) via persistent Digital Object Identifiers (DOIs).
By performing community detection (Louvain modularity) and structural hole analysis across multidisciplinary clusters (e.g., Edge TinyML, clinical Linked Data, and semantic RAG), the system algorithmically identifies topological voids—unexplored conceptual intersections with zero co-occurrence citations—actively directing researchers toward high-impact, unfilled research gaps.
Cite This Work
Select your preferred citation format to copy directly into your bibliography or manuscript:
S. F. Fajarudin, "CorpusLD: A Dual-Layer Semantic Extraction Framework and Deep Knowledge Graph Architecture for Scientific Literature with Deterministic Unit Ontology and Dynamic Authority Disambiguation," Zenodo, Aug. 2026. doi: 10.5281/zenodo.22179715.