Enhancing BERT-Based Chinese Address Recognition Model with Tag Revision Module
Ruonan Zhao, Xiangwu Ding · 2023
Address recognition refers to segmenting the address text and identifying the corresponding types. However, rule-based and conditional probability-based methods are limited in applicability and scalability, and deep learning-based methods may overlook features at the word level and exhibit lower accuracy than at the character level. To address these issues, this paper utilizes sequence labeling method and proposes a tag revision module to enhance Chinese address recognition based on the BERT pre-training model, which utilizes the high accuracy rate at the character level and features of address text. The experimental results on the Chinese address recognition dataset from the China Knowledge Graph and Semantic Computing Conference demonstrate the effectiveness of the proposed method, achieving an F1-score of 0.9402. Moreover, the proposed scheme in this paper improves the BERT model by 0.73 percentage points.