Scientific knowledge graph and ontology generation using open large language models | Alexandru Oarga et al. | ResearchPod

Transcript: Scientific knowledge graph and ontology generation using open large language models

Alex: Welcome to another episode of ResearchPod. Today we're discussing a new study that introduces a way to automatically turn massive amounts of scientific text into organized, searchable maps of knowledge.

Sam: So the core problem is that scientific fields are flooded with data, but most of it is just sitting in papers that computers can't really make sense of?

Alex: Exactly. Text is unstructured — it's written for human readers, not machines. A computer can store a paper, but it can't easily understand that one concept in that paper is related to a concept in a different paper published five years earlier. Most of that knowledge is effectively locked away.

Sam: And the traditional fix — having experts manually build a system to organize all those concepts — is just too slow for modern research?

Alex: It is. These organizing systems, called ontologies, can take years to build by hand. An ontology is essentially a formal rulebook for a field of knowledge. It defines what the important categories are, what counts as a member of each category, and how those categories relate to one another. Think of it like the Dewey Decimal System for a library, except instead of just sorting books into shelves, it also records that "mystery novels" and "thriller novels" are both types of "fiction," and that "fiction" is distinct from "biography." This study proposes an automated pipeline that uses AI to build these maps from scratch, without needing any hand-labeled training data.

Sam: How does the system figure out the relationships between different scientific ideas without being told what they are beforehand?

Alex: Think of it like an apprentice librarian who's handed a huge stack of specialized books and told to organize them. First, they read through everything and list every unique term they encounter. Then they start grouping those terms into broader topics. Finally, they draw a kind of family tree — scientists call this a "taxonomy" — showing how those topics nest inside one another. That family tree then becomes the skeleton for a "knowledge graph," which is a network where the dots are real-world concepts and the lines between them represent how those concepts connect.

Sam: So the taxonomy is the blueprint, and the knowledge graph is the actual map you can search and query?

Alex: That's a good way to put it. And the knowledge graph is what makes the information useful for AI systems. Instead of just retrieving a paragraph from a paper, a researcher can ask a question that requires reasoning across dozens of papers at once.

Sam: You mentioned this process is "zero-shot." What does that mean exactly?

In machine learning, "training" a model usually means showing it thousands of labeled examples so it learns the patterns. Zero-shot means skipping that step entirely. The system uses a large language model — the same kind of AI behind modern chatbots — and relies on its general reasoning ability to extract information directly from the text. It never sees a single hand-labeled example from this specific field. It just reads the papers and reasons from what it finds.

Scientific knowledge graph and ontology generation using open large language models

Alexandru Oarga, Matthew Hart, Andres M. Bran, Magdalena Lederbauer, Philippe Schwaller

Jun 17, 2026·5:57·user_upload

Knowledge Graphs (KGs)OntologyLarge Language Models (LLMs)Graph-Based Retrieval Augmented Generation (RAG)Single Atom Catalysts (SACs)Ontology Generation

Listen

6 min

Briefing

Research Approach

Knowledge graphs (KGs) are essential for structuring scientific data, but their creation is typically a labor-intensive task requiring domain experts to manually curate ontologies. The authors propose OntoGen, a five-step, zero-shot pipeline that automates this process using open-source large language models (LLMs). The pipeline extracts domain-specific vocabulary, identifies high-level categories, builds a hierarchical taxonomy, instantiates the KG, and finally extracts semantic relationships between entities. By using an incremental approach and self-consistency prompting, the system constructs a global, unified knowledge structure from a corpus of scientific papers.

Evaluation and Application

The authors evaluated OntoGen by reconstructing an existing chemical knowledge graph, demonstrating that it outperforms traditional extraction methods. To prove its utility in emerging fields where structured data is scarce, they applied the pipeline to Single Atom Catalysts (SACs). This resulted in the first domain-specific ontological schema and knowledge graph for SACs. The study also introduces a set of hierarchical metrics—hierarchical precision, recall, and accuracy—to evaluate the structural integrity of the generated taxonomies, moving beyond simple term-matching metrics.

Why It Matters

Automating the generation of ontologies and knowledge graphs significantly lowers the barrier to entry for building specialized scientific databases. By providing a structured foundation, these generated KGs can enhance the factuality and reasoning capabilities of LLMs through Graph-Based Retrieval Augmented Generation (GraphRAG). This work offers a scalable, interpretable, and efficient solution for managing knowledge in complex scientific domains, enabling researchers to extract actionable insights from vast amounts of unstructured literature without the need for expensive, domain-specific model training.

Key Terms and Definitions

Ontology — A formal conceptualization consisting of classes, relations, and instances that provides a structured schema for organizing knowledge.
Knowledge Graph (KG) — A representation of data as a network of nodes (entities) and edges (relationships) that facilitates structured information retrieval and reasoning.
Zero-Shot Learning — A machine learning paradigm where a model performs a task without having been explicitly trained on examples of that specific task.
Self-Consistency — A prompting strategy where a model generates multiple outputs for the same input, and the final result is determined by a majority vote to reduce hallucinations.
Single Atom Catalysts (SACs) — An emerging class of materials in catalysis where individual metal atoms are dispersed on a support, representing a complex scientific domain with limited pre-existing structured data.

Scientific knowledge graph and ontology generation using open large language models: Research Paper Podcast