Category–semantic fusion network for robust named entity recognition using context-aware representation learning

Named entity recognition (NER) is a subfield of NLP that aims to classify mentions of things such as people, places, and companies in unstructured text. Although transformer-based innovations have greatly enhanced contextual representation learning, the majority of current models still depend on semantic embeddings alone and lack explicit knowledge of entity-category data. In noisy or domain-diverse text, this constraint commonly causes boundary ambiguity, type misclassification, and unstable predictions. This research presents a Category-Semantic Fusion Network (CSFN) to overcome these obstacles; the network incorporates learnable category vectors with contextual word representations for robust NER. The suggested architecture includes an inference module for globally optimised sequence labelling using BiLSTM-CRF, as well as a Fusion Vector Generation module that integrates transformer-based embeddings with category-aware representations. Extensive experiments were carried out on three publicly available NER datasets, Weibo, Youku, and Chinese Literature, spanning the fields of social media, entertainment, and storytelling. According to the experimental results, the suggested model routinely outperforms the baseline BERT-BiLSTM-CRF architecture. On the Weibo dataset, the suggested framework improves the F1 score by 5.05%; on the Youku dataset, by 1.53%; and on the Chinese Literature dataset, by 1.81%. Additionally, the training stability study shows less variation among runs, indicating more reliable convergence. The results show that NER is considerably more accurate, resilient, and generalisable across different textual domains when category information is explicitly modelled alongside semantic context.