ArticleMiner: Ontology-Guided Knowledge Graph Construction from Scientific Publications

A hybrid framework combining parser evidence, language models, and bounded human-authored task modules for cross-domain scientific fact extraction

Md Abrar Jahin; Craig A. Knoblock; Jay Pujara· 2026· DOI 10.48550/arXiv.2609.25607

The core problem

Scientific publications embed much of their quantitative content in tables and supplementary files, where a number acquires meaning only through its header, caption, unit, analytical method, and field-specific conventions. Recovering the rows and columns of a table is therefore not equivalent to recovering the scientific fact it reports. Most semantic table-interpretation methods assume a clean table is already available and then map its cells or columns to ontology terms, while most publication-level extraction systems are designed for a single domain. This work studies a middle path: a shared process that reads a paper and its supplementary files, gathers evidence from several parsers and a language model, and reconciles that evidence, while a bounded human-authored task module for each task supplies the domain meaning. The module lists the canonical names the graph may use, the surface forms that map to them, a small set of derivation rules and validity constraints, an identity key, and the bindings used to write RDF. It defines what a task is allowed to emit; it does not try to list every convention of a field. Four such modules are built (for drug-discovery chemistry, material

ArticleMiner was evaluated on 163 papers across four domains: drug-discovery chemistry, materials science, machine learning, and mineral geochemistry. The geochemistry evaluation introduced a new benchmark with expert-curated ground truth. In comparisons against a same-LLM few-shot baseline, the point estimates favor ArticleMiner on all four tasks. For the two smaller benchmarks, the results carry uncertainty, but the direction of improvement is consistent. The geochemistry comparison is particularly informative because it includes access to supplementary files, so the observed improvement cannot be attributed to domain guidance alone. The framework successfully reconciled evidence from multiple parsers and a language model, and the bounded task modules effectively supplied domain meaning without attempting to enumerate every convention of a field. The new geochemistry benchmark provides a valuable resource for future work on scientific fact extraction.

Scientific publications embed much of their quantitative content in tables and supplementary files, where a number acquires meaning only through its header, caption, unit, analytical method, and field-specific conventions. Recovering the rows and columns of a table is therefore not equivalent to recovering the scientific fact it reports. Most semantic table-interpretation methods assume a clean table is already available and then map its cells or columns to ontology terms, while most publication-level extraction systems are designed for a single domain. This work studies a middle path: a shared process that reads a paper and its supplementary files, gathers evidence from several parsers and a language model, and reconciles that evidence, while a bounded human-authored task module for each task supplies the domain meaning. The module lists the canonical names the graph may use, the surface forms that map to them, a small set of derivation rules and validity constraints, an identity key, and the bindings used to write RDF. It defines what a task is allowed to emit; it does not try to list every convention of a field. Four such modules are built (for drug-discovery chemistry, materials science, machine learning, and mineral geochemistry) in the ArticleMiner framework, and evaluated on 163 papers, including a new geochemistry benchmark with expert-curated ground truth.

ArticleMiner follows a modular pipeline that separates evidence gathering from domain-specific interpretation. The architecture is illustrated below:

The results suggest that a shared extraction process combined with lightweight, human-authored task modules can effectively bridge the gap between generic table interpretation and domain-specific publication extraction. The bounded nature of the modules—listing canonical names, surface forms, derivation rules, validity constraints, identity keys, and RDF bindings—provides sufficient domain guidance while remaining maintainable and extensible. The improvement on the geochemistry task, which included supplementary files, indicates that the gains are not solely due to domain guidance but also stem from the reconciliation of multiple evidence sources. However, the uncertainty on the two smaller benchmarks highlights the need for larger-scale evaluations. Future work could expand the set of task modules, incorporate additional parser types, and explore automated module generation. The ArticleMiner framework offers a promising middle path for constructing knowledge graphs from scientific publications, balancing the generality of shared processes with the precision of domain-specific constraints.

CS practitioners and researchers

Opening member content…