Multi-Strategy Fusion for Medical Named Entity Recognition
Jian Sun, Rongchuan Tang, Lu Xiang, Feifei Zhai, Yu Ping Zhou · 2021
This paper describes a method of recognizing Chinese clinically named entities based on multi-strategy fusion. In this clinical medical entity recognition task, we need to identify the boundaries of 18 types of entities from the medical dialogue. There are some problems in this task, such as insufficient data, uneven distribution of entity types, and inconsistent annotation. We alleviate these problems through the integration of multiple strategies. We first use the semi-supervised learning method to expand the training data, and then use the ensemble learning method to integrate the recognition results of multiple models, reducing the recognition bias caused by a single model. Finally, dictionary matching, prefix matching, and other post-processing methods for fusion results, to maximize the solution of label inconsistencies. The F1 value of our proposed method on the final test set is 85.035801, ranking first.