Building knowledge graphs that invent their own schemas

What it does

AutoSchemaKG is a Python framework that converts unstructured text into knowledge graphs without a predefined ontology. It first uses LLMs to extract entities, events, and relations as triples, then induces a schema dynamically through conceptualization—grouping similar concepts to create semantic bridges across domains. The resulting graphs support multi-hop question answering and retrieval-augmented generation.

The interesting bit

Most KG construction tools treat schema design as a prerequisite. AutoSchemaKG treats it as a byproduct: the same LLM pipeline that extracts facts also generalizes them into higher-level categories. The authors used this autonomy to build ATLAS, a family of graphs with 900 million nodes and 5.9 billion edges drawn from Wikipedia, academic papers, and Common Crawl, with no manual schema curation.

Key highlights

• Distributed as the atlas-rag PyPI package with modules for triple extraction, vector storage, and retrieval
• Supports multilingual pipelines (Chinese, Japanese, Korean, etc.) and ingests PDF or Markdown via conversion tools
• Includes evaluation suites for KG quality, factual consistency on FELM, and general performance on MMLU
• Ships with Jupyter notebooks for hosting billion-scale graphs in Neo4j and benchmarking multi-hop QA on MuSiQue, HotpotQA, and 2WikiMultiHopQA
• Provides parallel generation scripts and custom prompt hooks for large-scale production use

Caveats

• The README claims state-of-the-art benchmark results but omits names, numbers, and comparisons, so the evidence is unclear
• Because extraction and conceptualization both rely on LLM inference, cost and latency scale linearly with corpus size
• The conceptualization step is described only at a high level; understanding the actual algorithm requires reading the paper

Verdict

A solid candidate if you want to bootstrap a domain-spanning knowledge graph from messy text without hand-crafting an ontology first. Less useful if you already have a fixed schema and just need fast, deterministic extraction.