ODE-CF: an Ordinary Differential Equation Cascaded Fusion Model for Chinese Named Entity Recognition
Haiyan Liu,Zhou Jin,Rujing Wang,Jie Ji,Zhiyuan Li,Yimin Hu
DOI: https://doi.org/10.1109/icnlp60986.2024.10692832
2024-01-01
Abstract:Chinese Named Entity Recognition (NER) is more complex than in other languages due to the fine granularity of Chinese words. Traditional Chinese NER works have focused on using large amounts of external knowledge to enhance entity detection. However, many professional fields lack sufficient data. To address these shortcomings, we propose an Ordinary Differential Equation Cascaded Fusion Model (ODE-CF) for Chinese NER. In contrast to Large Language Models (LLMs) that use text generation models, to take advantage of the embedding information ODE-CF uses a traditional sequence labeling task. At the same time, Compared to the single-tier structure, we employ the two-tier cascaded structure can obtain more Entity Position Information (EPI) from the first tier. Furthermore, We propose the ODE-Residual Fusion module to distill the EPI from the output of adjacent blocks in the first tier, which can be utilized in the second tier. Our experiments on the Chinese NER task demonstrate the effectiveness and genericity of this model. It achieves 96.41%, and 70.66% F1 values respectively on the MSRA corpus and Weibo corpus. In particular, further experiments on the small-scale Agris corpus, Youku corpus, and Ecommerce corpus show significant gains that 10.47%, 22.91% and 28.17% higher than the basemodel. It indicates our advantage in data-deficient professional fields, such as agriculture and finance.