Interactive Attention Network for Chinese Address Element Recognition

Abstract

Existing Named Entity Recognition (NER) models have achieved good performance, but they have low accuracy in Chinese address element recognition tasks. After analysis, we believe that the boundary information of the address text is more sensitive than the general text, and the sentences are independent of each other, unlike the general paragraph-style text with contextual connections. On the other hand, the previous NER models rarely consider the use of interaction between subtasks to enhance the performance of the NER task.This paper proposes an Interactive Attention Network (IAN) model, which uses boundary-based information and type-based information to improve NER task performance and introduces an interaction mechanism to share information between each subtask. In addition, a boundary auxiliary module is added to obtain explicit boundary information. The experimental results show that the proposed IAN model can solve the address element recognition task more effectively.