Title: Leveraging Structured Semantic Data to Enhance In-Context Learning for Few-Shot Relation Extraction

Abstract: This study explores methods for automatically generating supplementary examples for in-context learning, thereby shifting the paradigm of relation extraction from a 1-shot to a few-shot context. Central to our approach is a novel selection mechanism that prioritizes new instances based on the syntactic-semantic structural similarity to the initial 1-shot example. Our analysis reveals that this technique yields distinct word choices and sentence constructions when compared to examples generated by Large Language Models (LLMs). By integrating both approaches, the resulting hybrid system offers a more comprehensive understanding of the target relations than either method in isolation. The proposed framework demonstrates robust transferability across different datasets, specifically FS-TACRED and FS-FewRel, as well as across various LLM architectures, including Qwen and Gemma. Empirical results indicate that our hybrid model consistently surpasses alternative strategies, securing state-of-the-art performance on FS-TACRED and delivering significant improvements on a tailored subset of FewRel.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC