Automated Construction of Domain-Specific Knowledge Graphs
Opis
Domain-agnostic, Dockerised, and extendable pipeline for automated knowledge graph construction, evaluation, and filtering from any large text corpus. Users supply a domain corpus (PubMed abstracts, technical publications, regulatory documents, telecom transcripts, etc.); the service extracts subject—predicate—object triplets, scores them via a configurable multi-criteria framework, and delivers a curated Neo4j knowledge graph. The pipeline combines complementary extraction strategies, including rule-based open information extraction, supervised relation extraction frameworks, and prompting-based large language models. Document preprocessing includes parsing, deduplication, chunking, and metadata retention. LLM extraction uses structured prompting with role instructions, entity and relation guidelines, few-shot examples, and quality-focused instructions that prioritize relevance over completeness. Extracted triplets are then analyzed for volume, diversity, and redundancy, optionally filtered for near-duplicates, and evaluated with an LLM-as-a-judge framework over configurable criteria such as relevance, factual consistency, specificity, usefulness, evidence or grounding, and potential risk. Outputs can be exported in RDF, JSON-LD, or tabular formats and integrated with graph databases, semantic search, or RAG pipelines. The entire stack is delivered as Dockerised microservices via Docker Compose with GPU-accelerated inference.
Pričakovani rezultati
Users receive a domain-specific knowledge graph with quality-scored triplets, enabling filtering at user-defined quality thresholds. The curated knowledge graph can be used to ground LLMs for domain-specific question answering, power structured search and exploration interfaces, and provide reliable, traceable domain knowledge for research and decision support.
Kartica storitve
| Kategorija | AI development and deployment |
| Domena | Knowledge Engineering, Information Extraction |
| Vertikala | Cross-domain |
| Ciljni uporabniki | Researchers, Knowledge Engineers, Data Scientists, Organisations Building AI-Powered Search/QA/Decision-Support Systems, Public-Administration Bodies, Telecom and Digital-Society Companies, SMEs and Start-ups |
| Tip storitve | Pipeline, Model, Data |
| Rezultat | Neo4j knowledge graph database with RDF/JSON-LD/tabular exports; GraphRAG-ready knowledge graph for retrieval-augmented generation; quality-scored triplets with per-criterion and composite scores; domain-adapted text corpus with metadata; surrogate scoring models; LLM-as-a-judge evaluation framework; filtering and export tools. Entire stack delivered as Docker containers. |
| Zahteve | Machine-readable text corpus in any target domain. Optional schema, ontology, or relation inventory for constrained extraction. NVIDIA GPU node recommended for local LLM inference and large-scale batch processing, although preprocessing and baseline extractors can run on CPU infrastructure. Optional domain expert input is useful for calibration of evaluation criteria, threshold selection, and validation of high-value triplets. |
| Omejitve | Quality depends on the coverage, cleanliness, and representativeness of the source corpus. Automated evaluation provides a scalable proxy for triplet quality, but expert validation is recommended for high-stakes settings. Domain adaptation may require custom entity types, relation definitions, ontologies, and evaluation criteria. |
| Na voljo od | Q2/2027 |
| Licenca | Open (Apache 2.0 for pipeline code; model weights under respective base model licences; CC BY 4.0 for evaluation framework) |
| Slika | |
| Oznaka slike |