Year: 2025
Type: article
Abstract: As pivotal geospatial information, address data requires robust parsing to effectively support critical internet-based services, including navigation and e-commerce logistics. Yet, existing deep learning-based Chinese address parsing is still confronted by challenges arising from data annotation scarcity, administrative redistricting, regional variations, and informal expressions. Thus, this study enhances model generalization by leveraging contextual prior knowledge while fundamentally reforming the conventional paradigm that frames address parsing as a sequence labeling task. A Gazetteer- and Type-enhanced BERT with an efficient global pointer (GTBERT-Pointer) is proposed, representing an innovative architecture integrating a pre-trained BERT encoder, plug-and-play gazetteer adapter, type adapter, and enhanced global pointer mechanism. The BERT layer first derives the contextual vector representations of the address sequences. Then, the novel gazetteer and type adapters strategically infuse external gazetteer resources and hierarchical element-type knowledge into deep contextual embeddings. Finally, address parsing is recast as multi-label classification through the efficient global pointer to achieve global-span extraction of element boundaries. Empirical evaluations on recent Chinese address parsing benchmarks demonstrate the state-of-the-art performance of GTBERT-Pointer, surpassing existing approaches across comprehensive metrics.
Authors: Jiachen Huang, Shuo Liu, Yanyou Qiao, Yuang Wu
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,390
Award: G660200874
OpenAlex ID: W4417037956