Our Publications

The standard evaluation metrics for DyGIE++ often employ a strict matching criterion that can underestimate model performance by penalizing partially correct extractions common in complex scientific text. This work introduces a relaxed matching strategy with a configurable threshold for accepting predictions as true positives. By applying this method to a curated fuel cell dataset, we demonstrate a significant and more realistic improvement in F1-scores for both Named Entity Recognition (NER) and Relation Extraction (RE), providing a more accurate assessment of the model's true capabilities in materials informatics.

This paper presents a novel framework for knowledge discovery in fuel cell research by integrating advanced Natural Language Processing (NLP) with Retrieval-Augmented Generation (RAG). We demonstrate a system that builds a vectorized knowledge base from thousands of scientific articles and allows researchers to query this base using natural language questions. Our results show that the RAG-enhanced Large Language Model (LLM) provides more accurate, context-aware, and verifiable answers compared to a baseline LLM, effectively mitigating hallucinations and enabling rapid synthesis of domain-specific information.