The Answer Lies Within: Self-Derived Rewards Enable Explainable Relation Extraction

The paper presents a novel framework for one-shot relation extraction that integrates COGRE, a cognitively-inspired reasoning structure, and HIT@DICT, a reinforcement learning strategy for deriving self-rewards from a credit dictionary. Experiments demonstrate that this approach, particularly with the Qwen2.5-14B-Instruct model, achieves a 24.65% F1 score on the NYT29 dataset and improves performance by 23.46% through reinforcement learning. This framework enhances both accuracy and explanation quality, providing practitioners with a method to align model outputs more closely with human expectations in relation extraction tasks.