SumPrompt: Summary-guided prompt tuning for few-shot relation extraction

Few-shot Relation Extraction (FSRE) aims to train effective Relation Extraction (RE) models with limited samples to identify semantic relations between entity pairs in text. In recent years, prompt-based approaches have emerged as a research hotspot in FSRE for their ability to leverage the potential of pre-trained language models. However, existing models generally overlook the interference caused by noisy information in the original text. To address this issue, this paper proposes a Summary-guided Prompt tuning model (SumPrompt). Specifically, the model employs a two-stage prompt to guide the Large Language Model (LLM) in generating and refining sample summaries without additional training, focusing on core entities and relations. Furthermore, with prompt categorization, it dynamically constructs summary prompts for a Small Language Model (SLM) based on the LLM's outputs. Experiments on TACRED, TACREV, and Wiki80 demonstrate that SumPrompt achieves state-of-the-art performance in most few-shot settings. Notably, under the K=1 setting, the F₁ score exceeds the best SLM-based approach by 2.8%, 3.3%, and 2.3% on the three datasets, respectively. This result not only demonstrates the effectiveness of the summary-guided prompt but also provides a new perspective for FSRE, as enhancing key information through text compression can significantly improve model robustness and generalization in noisy environments.