多模态文本信息的高速公路交通事件持续时间预测

Prediction of traffic incident duration on expressway based on multimodal text information

为研究文本信息质量对高速公路应急处置的影响,针对多阶段自然语言描述的交通事件文本信息,提出基于预训练模型和深度学习的多模态数据持续时间预测架构,通过对比分析Word2vec、BERT、ALBERT和RoBERTa预训练模型在BiLSTM-CNN基准预测模型上的性能指标,检验预训练后词向量在原始文本、首次文本和拼接文本3种模态的准确性和稳健性。研究结果表明:Word2vec-BiLSTM-CNN模型预测性能较优,训练时间短且具有较好的鲁棒性,能够适用于不同模态的文本数据;相较于原始文本和拼接文本,首次文本信息预测效果良好,是交通事件持续时间预测的关键;拼接文本虽然增加了训练时间,但预测性能并未得到提高。研究结果可有效论证文本信息质量对高速公路应急处置的重要作用,为后续多阶段文本信息数据处理提供支撑。

In order to study the influence of the text information quality on the expressway emergency response,a prediction architecture of multimodal data duration based on pre-training model and deep learning was proposed aiming at the text information of traffic incident described by multi-stage natural language.By comparing and analyzing the performance indexes of Word2vec,BERT,ALBERT and RoBERTa pre-training models on the BiLSTM-CNN benchmark prediction model,the accuracy and robustness of the pre-training word vector in the three modes of original text,first text and spliced text were tested.The results show that Word2vec-BiLSTM-CNN model has better prediction performance,short training time and good robustness,and can be applied to text data with different modes.Compared with the original text and spliced text,the prediction effect of first text information is good,which is the key to the prediction of traffic incident duration.The spliced text increases the training time,but do not improve the prediction performance.The research results can effectively demonstrate the important role of text information quality in expressway emergency response,and provide support for subsequent multi-stage text information data processing.