Embedding-based Entity Expansion for Data Augmentation in Materials Science

NLP-based techniques have demonstrated notable success in tasks like entity recognition within materials science [1] [2]. However, the field suffers from data scarcity due to the resource-intensive nature of generating high-quality annotated data. To address this issue, data augmentation is one of the promising approaches. Yet, prior studies [3] [4] primarily transform human-annotated entities, which are often noisy and incomplete, potentially limiting the diversity needed for robust model development. To address this, we propose EXPAND, an embedding-based entity expansion framework for data augmentation in the materials science domain. EXPAND leverages materials-aware embeddings derived from ChemDataExtractor [5] and Mat2Vec [6] to expand the range of entities beyond human annotations. Our proposed method achieves a 1.5% improvement over baseline methods across various materials science benchmarks, demonstrating its effectiveness in overcoming data scarcity.