Semi-supervised named entity recognition in low-resource domains: A case study of rare earth elements in coal

Abstract

Geological literature serves as a crucial repository of information for advancing geological understanding and guiding mineral exploration. However, manually extracting geological information from unstructured text is labor-intensive and inefficient. Named entity recognition (NER), a core task in information extraction (IE), offers an automated solution by identifying and classifying geological entities. While NER has been widely applied in geological fields, it remains limited in coal geology and coal-hosted critical metal deposits, which represent a typical low-resource domain. Therefore, this work proposes an efficient approach for the low-resource NER, using rare earth elements (REE) in coal as a representative case. An unlabeled text corpus is first collected from the Web of Science (WoS) database. Distant supervision and large language models (LLM) are then integrated to directly annotate part of the corpus. Domain experts subsequently validate and refine these annotations, yielding a high-quality labeled dataset for training the semi-supervised NER model. Specifically, four widely used NER models (i.e., BERT, BERT-CRF, BiLSTM-CRF, and BERT-BiLSTM-CRF models) are compared to identify the optimal model for self-training. This work focuses on seven geological entity types, including rock, mineral, element, maceral, location, stratum, and geologic time. Experimental results show that the BERT-CRF model achieves the best performance, with an F1-score of 0.8467. After applying the self-training algorithm, the F1-score improves to 0.8702, highlighting the effectiveness of the proposed approach in enhancing NER performance in the low-resource domain. Additionally, an entity-based literature retrieval system is developed to facilitate the accurate and efficient extraction of geological information from relevant sources.