The Answer Lies Within: Self-Derived Rewards Enable Explainable Relation Extraction
Researchers have developed a new framework called COGRE that enhances the explainability and accuracy of relation extraction in large language models. This framework addresses challenges such as models being misled by irrelevant text and failing to match human annotator expectations. COGRE structures the extraction process to mimic human text processing and uses a reinforcement learning strategy, HIT@DICT, to align reasoning with relational labels by rewarding relation-relevant phrases derived from correct predictions. AI
IMPACT Introduces a novel approach to improve LLM performance and interpretability in relation extraction tasks.