Imbalanced Traffic Accident Text Classification Based on Bert-RCNN
Shijin Yuan, Qixuan Wang
Abstract
Abstract Traffic is the signs of urban development, and traffic accidents have a particularly large impact on road conditions. Because traffic accidents are recorded in text, classifying the severity of traffic accidents can not only unify the classification standards, but also greatly help query and prediction of traffic accidents. The frequency of daily minor traffic accidents is much higher than that of serious traffic accidents, so this paper proposes an effective algorithm and model for categorizing imbalanced Chinese traffic accident texts based on severity. Firstly, we establish a standard for refined classification granularity combining expert knowledge, and propose unique accuracy standards to verify the impact of the processed dataset on the model. Then the noise on the dataset is removed, and finally input the text into Bidirectional Encoder Representations from Transformers (BERT) model uses the Recurrent Convolutional Neural Network (RCNN) model as a classifier for fine-tuning. By comparing with the results of other classification methods, it is verified that our algorithm and model have the best accuracy.