Open Access

ARTICLE

Domain-Specific NER for Fluorinated Materials: A Hybrid Approach with Adversarial Training and Dynamic Contextual Embeddings

Jiming Lan¹, Hongwei Fu^1,*, Yadong Wu^1,2, Yaxian Liu^1,3, Jianhua Dong^1,2, Wei Liu^1,2, Huaqiang Chen^1,2

1 School of Computer Science and Engineering, Sichuan University of Science and Engineering, Zigong, 644005, China
2 Sichuan Engineering Research Center for Big Data Visual Analytics, Zigong, 644005, China
3 Key Laboratory of Higher Education of Sichuan Province for Enterprise Informationalization and Internet of Things, Sichuan University of Science and Engineering, Zigong, 644005, China

* Corresponding Author: Hongwei Fu. Email:

Computers, Materials & Continua 2025, 85(3), 4645-4665. https://doi.org/10.32604/cmc.2025.067289

Received 29 April 2025; Accepted 22 July 2025; Issue published 23 October 2025

Abstract

In the research and production of fluorinated materials, large volumes of unstructured textual data are generated, characterized by high heterogeneity and fragmentation. These issues hinder systematic knowledge integration and efficient utilization. Constructing a knowledge graph for fluorinated materials processing is essential for enabling structured knowledge management and intelligent applications. Among its core components, Named Entity Recognition (NER) plays an essential role, as its accuracy directly impacts relation extraction and semantic modeling, which ultimately affects the knowledge graph construction for fluorinated materials. However, NER in this domain faces challenges such as fuzzy entity boundaries, inconsistent terminology, and a lack of high-quality annotated corpora. To address these problems, (i) We first construct a domain-specific NER dataset by combining manual annotation with an improved Easy Data Augmentation (EDA) strategy; (ii) Secondly, we propose a novel model, RRC-ADV, which integrates RoBERTa-wwm for dynamic contextual word representation, adversarial training to improve robustness against boundary ambiguity, and a Residual BiLSTM (ResBiLSTM) to enhance sequential feature modeling. Further, a Conditional Random Field (CRF) layer is incorporated for globally optimized label prediction. Experimental results demonstrate that RRC-ADV achieves an average F1 score of 89.23% on the self-constructed dataset, significantly outperforming baseline models. The model exhibits strong robustness and adaptability within the domain of fluorinated materials. Our work enhances the accuracy of NER in the fluorinated materials processing domain and paves the way for downstream tasks such as relation extraction in knowledge graph construction.

---

Keywords

---