Abstract

Scientific knowledge is still communicated primarily through papers, while the actual objects that science accumulates are more granular: claims, methods, datasets, experimental traces, replications, contradictions, and theories. This mismatch has become more severe as scientific literature grows rapidly and large language models (LLMs) make the production and consumption of scientific text easier but not necessarily more reliable. In this blue-sky paper, we argue that semantic technologies should treat scientific assertions as first-class objects and support the creation of evidence-centric scientific knowledge graphs (KGs) directly from literature. We call this vision semantics for science. We argue that building such KGs is not simply a larger information extraction problem. It requires explicit representations of complex evidence traces, including provenance, uncertainty, scope conditions, temporal evolution, evidential support, and challenge relations. We further argue that the resulting infrastructure may require a reasoning engine for science that goes beyond conventional description logics entailment or graph traversal, combining symbolic semantics with defeasible, provenance-aware, and evidence-weighted reasoning. We outline a research agenda spanning representation languages, extraction pipelines, identity and alignment, contradiction handling, temporal revision, and benchmark design. Our claim is deliberately ambitious: if the Semantic Web community wants to shape how scientific knowledge is represented in the age of language models and AI agents, it should help move science from paper-centric publication systems toward …

Date

2026

Authors

Mayank Kejriwal