Character-level CNN for structural parsing and named entity recognition of museum costume terminology

Abstract

To address the unstructured and semantically complex nature of cataloging terminology for museum costume artifacts, this paper proposes a Character-Level Convolutional Neural Network (CNN) for Named Entity Recognition (NER) to achieve automated structural analysis. This study formalizes the problem as a sequence labeling task using the BIO tagging scheme and develops an end-to-end CNN model based on character embeddings. By employing parallel multi-size convolutional filters, the model automatically learns combinatorial features from raw character sequences without reliance on manually crafted rules. Experimental results demonstrate that the proposed model achieves a strong macro-average F1-score on the manually annotated validation set, validating its effectiveness for the structural analysis of costume terminology. This method lays a technical foundation for building domain-specific knowledge graphs and enhancing intelligent retrieval in cultural heritage systems, facilitating advanced semantic queries such as cross-dynasty pattern comparisons and technical classification.