Enhancing Named Entity Recognition through Advanced Sequence Labeling using Conditional Random Fields with Contextual Feature Integration

Abstract

Objectives:  This study aims to improve the accuracy and consistency of Named Entity Recognition (NER) systems by addressing the limitations of transformer-based models that rely on independent token-level predictions. It focuses on exploring how structured prediction techniques can enhance sequence labeling performance. Methods: A comparative evaluation is carried out using three hybrid architectures: BERT combined with CRF, BiLSTM combined with CRF, and BERT integrated with a Transformer layer. All models are tested within a consistent experimental setup using the CoNLL-2003 dataset. The methodology emphasizes combining contextual representations with sequence-level optimization through Conditional Random Fields. Findings: The results indicate that the BERT + CRF model delivers the highest performance, achieving an F1-score of 0.94. In comparison, BERT + Transformer attains 0.81, while BiLSTM + CRF reaches 0.72. These outcomes demonstrate that incorporating CRF significantly improves label sequence coherence and overall extraction effectiveness. Novelty: The work presents a unified and systematic comparison of hybrid NER models, highlighting the importance of structured decoding in modern architectures. It offers valuable insights into balancing model performance and computational efficiency, supporting the development of reliable and scalable NER solutions for practical applications. The other NER models often fail to classify multi-token organization entities and entities with ambiguous contextual meaning correctly. The BERT + CRF model got better boundary detection by structured sequence optimization.