[KongchangAI]
BenchmarkSemantic Structural Similarity for KGs / 知识图谱语义结构相似度

S3KG

A knowledge-graph-based evaluation framework for assessing LLM contextual understanding that integrates both structural and semantic similarity signals. It evaluates the factual accuracy and logical consistency of model responses through triple-level analysis.

Timeline (last 90 days)

Sep 28

The research team proposed a knowledge-graph-based evaluation framework whose core is a hybrid similarity measure called S3KG (Semantic Structural Similarity for KGs).

Unverified50%
Sep 28

S3KG captures both structural and semantic signals simultaneously: structural similarity measures whether the knowledge structures align, semantic similarity ensures that entities and relations are consistent at the meaning level, and the two signals are fused into a single score.

Unverified50%
Sep 28

The study also developed a companion diagnostic analysis framework capable of identifying and categorizing reasoning errors at the triple level, enabling fine-grained analysis of model failure cases.

Unverified50%
Sep 28

The research team validated the effectiveness of S3KG on nine benchmarks.

Unverified50%
Sep 28

Compared to the strongest baseline method, S3KG achieves an F1 score improvement of up to 7.6 percentage points.

Unverified50%
Sep 28

S3KG achieves a peak AUROC of 0.973.

Unverified50%

All Facts (6)

Source Articles