KICE: A Knowledge Consolidation and Expansion Framework for Relation Extraction

Machine Learning is often challenged by insufficient labeled data. Previous methods employing implicit commonsense knowledge of pre-trained language models (PLMs) or pattern-based symbolic have achieved great success in mitigating manual annotation efforts. In this paper, we focus on the collaboration among different sources and present KICE, a Knowledge-evolving framework Iterative Consolidation Expansion with guidance PLMs rule-based patterns. Specifically, starting limited data as seeds, KICE first builds Rule Generator prompt-tuning to stimulate rich distributed PLMs, generate seed rules, initialize rules set. Afterwards, based rule-labeled data, task model trained self-training pipeline where set consolidated self-learned high-confidence rules. Finally, for low-confidence solicits human-enlightened understanding expands coverage better training. Our verified relation extraction (RE) task, experiments TACRED show that performance (F1) grows from 33.24% 79.84% enrichment knowledge, outperforming all baselines including other knowledgeable methods.