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

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot Relation Extraction (FSRE) models require effective training with limited data.
• Prompt-based approaches leverage pre-trained language models but often struggle with noisy text.
• Existing methods overlook the impact of irrelevant information on relation extraction performance.

Purpose of the Study:

• To propose a Summary-guided Prompt tuning model (SumPrompt) to address noisy information in FSRE.
• To enhance the robustness and generalization of relation extraction models in low-data scenarios.
• To improve the performance of models by focusing on core entities and relations through text compression.

Main Methods:

• A two-stage prompt guides a Large Language Model (LLM) to generate and refine sample summaries without additional training.
• Prompt categorization dynamically constructs summary prompts for a Small Language Model (SLM) based on LLM outputs.
• The SumPrompt model focuses on core entities and relations within the generated summaries.

Main Results:

• SumPrompt achieves state-of-the-art performance across multiple few-shot settings on TACRED, TACREV, and Wiki80 datasets.
• At K=1, SumPrompt's F1 score surpasses the best SLM-based approach by 2.8% (TACRED), 3.3% (TACREV), and 2.3% (Wiki80).
• The summary-guided prompt significantly improves model robustness and generalization in noisy environments.

Conclusions:

• The proposed SumPrompt model effectively enhances few-shot relation extraction performance.
• Text compression via summarization is a viable strategy to improve model robustness and generalization.
• This work offers a new perspective for FSRE by leveraging text summarization to mitigate noise.