Paper ID: 2312.11539

KGLens: A Parameterized Knowledge Graph Solution to Assess What an LLM Does and Doesn't Know

Shangshang Zheng, He Bai, Yizhe Zhang, Yi Su, Xiaochuan Niu, Navdeep Jaitly

Measuring the alignment between a Knowledge Graph (KG) and Large Language Models (LLMs) is an effective method to assess the factualness and identify the knowledge blind spots of LLMs. However, this approach encounters two primary challenges including the translation of KGs into natural language and the efficient evaluation of these extensive and complex structures. In this paper, we present KGLens--a novel framework aimed at measuring the alignment between KGs and LLMs, and pinpointing the LLMs' knowledge deficiencies relative to KGs. KGLens features a graph-guided question generator for converting KGs into natural language, along with a carefully designed sampling strategy based on parameterized KG structure to expedite KG traversal. We conducted experiments using three domain-specific KGs from Wikidata, which comprise over 19,000 edges, 700 relations, and 21,000 entities. Our analysis across eight LLMs reveals that KGLens not only evaluates the factual accuracy of LLMs more rapidly but also delivers in-depth analyses on topics, temporal dynamics, and relationships. Furthermore, human evaluation results indicate that KGLens can assess LLMs with a level of accuracy nearly equivalent to that of human annotators, achieving 95.7% of the accuracy rate.

Submitted: Dec 15, 2023