OntoDoc
A multi-agent framework for automatically generating natural language descriptions of resources in a knowledge graph, combining the internal structure of the knowledge graph with external encyclopedic knowledge.
Timeline (last 90 days)
OntoDoc is a multi-agent framework that can automatically generate natural language descriptions for any Internationalized Resource Identifier (IRI) in a SPARQL endpoint or RDF file.
OntoDoc's workflow is divided into four stages, each handled by a dedicated agent: the database context agent, the web context agent, the hypothesis generation agent, and the description generation agent.
The database context agent adaptively explores the knowledge graph via SPARQL, dynamically constructing queries according to the type of the target IRI (class, property, or instance) and collecting triples semantically related to the target IRI.
The web context agent uses the local name of the IRI or a label extracted from the graph as query terms, retrieves encyclopedic background knowledge from Wikipedia and the open web, and reranks the retrieval results.
The hypothesis generation agent integrates the database context and the web context to synthesize a structured list of evidence-supported intermediate hypotheses.
The description generation agent generates the final natural language description based on the grounded hypotheses, which have already undergone evidence verification.
The researchers evaluated OntoDoc on three knowledge graphs, covering biomedical, agricultural, and cross-domain scenarios respectively.
The evaluation adopted two approaches: automatic scoring based on large language models (LLMs) and human evaluation by 11 annotators.
OntoDoc consistently meets or exceeds naive LLM baselines and template-based ontology verbalizers in accuracy.
In terms of completeness, OntoDoc improves by an average of 12% over the naive baseline across all domains, and improves completeness scores by up to 2.8 points on a 5-point scale relative to ontology verbalizers.
1 more timeline events
All Facts (11)
OntoDoc is a multi-agent framework that can automatically generate natural language descriptions for any Internationalized Resource Identifier (IRI) in a SPARQL endpoint or RDF file.
50%UnverifiedOntoDoc's workflow is divided into four stages, each handled by a dedicated agent: the database context agent, the web context agent, the hypothesis generation agent, and the description generation agent.
50%UnverifiedThe database context agent adaptively explores the knowledge graph via SPARQL, dynamically constructing queries according to the type of the target IRI (class, property, or instance) and collecting triples semantically related to the target IRI.
50%UnverifiedThe web context agent uses the local name of the IRI or a label extracted from the graph as query terms, retrieves encyclopedic background knowledge from Wikipedia and the open web, and reranks the retrieval results.
50%UnverifiedThe hypothesis generation agent integrates the database context and the web context to synthesize a structured list of evidence-supported intermediate hypotheses.
50%UnverifiedThe description generation agent generates the final natural language description based on the grounded hypotheses, which have already undergone evidence verification.
50%UnverifiedThe researchers evaluated OntoDoc on three knowledge graphs, covering biomedical, agricultural, and cross-domain scenarios respectively.
50%UnverifiedThe evaluation adopted two approaches: automatic scoring based on large language models (LLMs) and human evaluation by 11 annotators.
50%UnverifiedOntoDoc consistently meets or exceeds naive LLM baselines and template-based ontology verbalizers in accuracy.
50%UnverifiedIn terms of completeness, OntoDoc improves by an average of 12% over the naive baseline across all domains, and improves completeness scores by up to 2.8 points on a 5-point scale relative to ontology verbalizers.
50%UnverifiedOntoDoc's method still depends on the quality of external web knowledge sources, and the evaluation scale is limited.
50%