Conceptgraph-grounded reference-free LLM-based evaluation / 图接地的无参考LLM评估
Graph-Grounded Evaluation
A reference-free evaluation method that bypasses reference texts and anchors directly on the input knowledge graph, in which an LLM judges the factual consistency between generated text and the input graph and can detect factual errors such as triple omission and contradiction.
Timeline (last 90 days)
Oct 23
For KG-to-text tasks, graph-grounded evaluation should become standard practice; relying solely on reference metrics such as BLEU and ROUGE may underestimate LLM capabilities and overlook the factual deficiencies of fine-tuned models.
Unverified50%