Active Knowledge Structuring for Large Language Models in Materials Science Text Mining

Large Language Models (LLMs) offer apromising alternative to traditional MaterialsScience Text Mining (MSTM) by reducingthe need for extensive data labeling andfine-tuning. However, existing zero-/few-shotmethods still face limitations in aligning withpersonalized needs in scientific discovery. Toaddress this, we propose ClassMATe, an activeknowledge structuring approach for MSTM.Specifically, we first propose a class defini-tion stylization method to structure knowledge,enabling explicit clustering of latent materialknowledge in LLMs for enhanced inference.To align with the scientists' needs, we proposean active needs refining strategy that iterativelyclarifies needs by learning from uncertainty-aware hard samples of LLMs, further refiningthe knowledge structuring. Extensive experi-ments on seven tasks and eight datasets showthat ClassMATe, as a plug-and-play method,achieves performance comparable to super-vised learning without requiring fine-tuningor extra knowledge base, highlighting thepotential to bridge the gap between LLMs'latent knowledge and real-world scientificapplications.