Junwoo Kim

Current entity alignment benchmarks are too easy; this paper introduces a hard-negative dataset where models that rely on name overlap fail spectacularly, but a new LLM-enhanced method succeeds.