Enhancing Legal Named Entity Recognition Using RoBERTa-GCN with CRF: A Nuanced Approach for Fine-Grained Entity Recognition

Accurate identification of named entities is pivotal for the advancement of sophisticated legal Artificial Intelligence (AI) applications. However, the legal domain presents distinct challenges due to the presence of fine-grained, domain-specific entities, including lawyers, judges, courts, and precedents. This necessitates a nuanced approach to Named Entity Recognition (NER). In this paper, we introduce a novel NER approach tailored to the legal domain. Our system combines Robustly Optimized BERT (RoBERTa) with a Graph Convolutional Network (GCN) to harness two distinct types of complementary information related to words in the data. Furthermore, the application of a Conditional Random Field (CRF) at the output layer ensures global consistency in data labeling by considering the entire sequence when predicting a named entity. RoBERTa captures contextual information about individual words, while GCN allows us to exploit the mutual relationships between words, resulting in more precise named entity identification. Our results indicate that RoBERTa-GCN (CRF) outperforms other standard settings, such as, RoBERTa, textGCN, and BiLSTM, including state-of-the-art for NER in the legal domain.