Boosting Triple Extraction and Relation Normalization with Ensembled Lightweight LLMs

Accurate and efficient construction of knowledge graphs heavily relies on the extraction and normalization of entity-relation triples from unstructured text. While large language models (LLMs) have demonstrated strong capabilities in these tasks, their high computational cost and deployment barriers limit their practical use in resource-constrained environments. In this paper, we propose a novel framework that leverages ensembled lightweight LLMs to boost both triple extraction and relation normalization performance. Our approach integrates multiple distilled or quantized LLMs through a consensus-based ensemble strategy, where each model contributes complementary strengths in identifying entities, recognizing relations, and resolving semantic ambiguities. To enhance robustness, we further introduce negative-example prompting to guide models away from common error patterns such as boundary misalignment and relation type confusion. For relation normalization, we present a definition-aware alignment method that utilizes semantic definitions generated by LLMs to compute fine-grained similarity between heterogeneous relation names. This enables more accurate mapping across different syntactic expressions of the same underlying semantics. Experimental results on benchmark datasets show that our ensembled lightweight LLM framework achieves state-of-the-art performance across all three evaluation metrics: Partial, Strict, and Exact, outperforming existing methods in both extraction accuracy and normalization effectiveness.