Case Study
SarnSapt - Thai Lexical Knowledge Graph & Grounded QA
A Thai lexical knowledge graph (17.2M triples) with a graph-grounded QA pipeline - Top 10 team, Dictionary Reimagined Hackathon 2026.
- Timeline
- Aug 2026 - Sep 2026
- Stack
- PythonGraphDBRDFSPARQLOntoLex-LemonSKOSPROVFastAPINext.jsThaiLLM
- Links
- RDF triples
- 17.2MRDF triples
- word senses from 13 datasets
- 272K+word senses from 13 datasets
- dialect-word accuracy
- 0 → 92%dialect-word accuracy
- verifiable source citations
- 100%verifiable source citations

01 / Problem
Thai lexical data is spread across many dictionaries, editions, and datasets. General-purpose LLMs answer word-meaning questions confidently but without sources - they miss dialect words and technical terms, and will even define words that do not exist.
02 / Approach
Engineered a provenance-preserving lexical knowledge graph in GraphDB using RDF, OntoLex-Lemon, SKOS, and PROV, integrating 13 datasets while keeping each source and dictionary edition in its own named graph. On top of it, built a graph-grounded question-answering pipeline with FastAPI, SPARQL, and ThaiLLM: context-aware sense selection, backend validation of every cited sense/evidence ID, source traceability, and graph-based fallback answers when the LLM is unavailable.
03 / Outcome
Selected as a Top 10 Qualified Team in the “เปิดคลังคำ พลิกคลังคิด” (Dictionary Reimagined) Hackathon 2026 by the Office of the Royal Society. On a 60-word stratified benchmark using the same Pathumma-ThaiLLM-qwen3-8b model, the system improved on the model-only baseline from 0% → 92% on dialect-word reference match, 33% → 100% on technical-term reference match, 92% → 100% on fabricated-word refusal, and 0% → 100% on verifiable source citation.