Chinese Fine‐Grained Geological Named Entity Recognition With Rules and FLAT. Issue 12 (28th November 2022)

• Record Type: Journal Article
• Authors: Chen, Siying; Hua, Weihua; Liu, Xiuguo; Deng, Xiaotong; Zeng, Xinling; Duan, Jianchao
• Is Part Of: Earth and space science. Volume 9:Issue 12(2022)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-11-28
• Subjects: geological text -- named entity recognition -- fine‐grained -- rules -- transformer; Space sciences -- Periodicals; Geophysics -- Periodicals; 500.5
• DOI: 10.1029/2022EA002617
• Languages: English
• ISSNs: 2333-5084

Abstract

Abstract: Geological named entity recognition (NER) is an essential prerequisite to realizing geological information extraction and information retrieval and is an actual means for accomplishing structured reconstruction of unstructured geological data. Existing geological NER methods mainly focus on coarse‐grained geological entity recognition, but geological entities are fine‐grained. To solve this problem, a Chinese fine‐grained geological entity corpus encompassing 21 types of fine‐grained labels is constructed. In addition, in this article, a fine‐grained geological entity recognition model based on Bidirectional Encoder Representations from Transformer (BERT)‐Flat‐Lattice Transformer is designed. This paper names this method FGNER (F ine‐grained G eological N amed E ntity R ecognition) which adds geological naming rules to revise the model results to improve the recognition of complex geological entities. The fine‐grained geological entity recognition method is evaluated using regional geological literature reports as experimental data. The experimental results show that the precision, recall, and F1‐score of the FGNER model are 95.73%, 89.26%, and 92.05%, respectively, thus achieving better performance than baseline models, such as BERT‐Conditional Random Field. Plain Language Summary: Most of the past geological named entity recognition (NER) methods have focused on studies of rocks and geological formations, but geologically named entities are fine‐grained. In this paper, we build a fine‐grained corpus and use a fusion of deep learning and rules to improve the recognition performance. We hope that this research will provide new priorities to advance the progress of geologically NER. Key Points: Fine‐grained entity categories are focused upon, in particular extracting 21 types of geological entities from geological literature Deep learning and rules are fused to extract geological entities from geological texts A fine‐grained corpus of geological entities is constructed based on a geological domain‐specific lexicon