交通运输系统工程与信息

交通运输系统工程与信息 ›› 2026, Vol. 26 ›› Issue (5): 122-131.DOI: 10.16097/j.cnki.1009-6744.2026.05.010

• 综合交通运输体系 • 上一篇    下一篇

基于社交媒体数据的交通突发事件提取方法

陈坚*1a ，胡功明1b ，简洪波2 ，傅志妍3   

  1. 1. 重庆交通大学，a.智慧城市学院，b.交通运输学院，重庆400074；2.陆海新通道运营有限公司，重庆401333； 3. 重庆第二师范学院，经济与工商管理学院，重庆400067
  • 收稿日期:2026-03-27 修回日期:2026-06-03 接受日期:2026-06-22 出版日期:2026-10-25 发布日期:2026-10-21
  • 作者简介:陈坚（1985—），男，江西赣州人，教授，博士。
  • 基金资助:
    重庆市社会科学规划项目 (2025BS059)；国家自然科学基金面上项目(52472339)。

Traffic Emergency Event Extraction Method Based on Social Media Data

CHEN Jian*1a, HU Gongming1b, JIAN Hongbo2, FU Zhiyan3   

  1. 1a. College of Smart City, 1b. College of Traffic and Transportation, Chongqing Jiaotong University, Chongqing 400074, China; 2. New Land-Sea Corridor Operation Co Ltd, Chongqing 401333, China; 3. School of Economics and Business Administration, Chongqing University of Education, Chongqing 400067, China
  • Received:2026-03-27 Revised:2026-06-03 Accepted:2026-06-22 Online:2026-10-25 Published:2026-10-21
  • Supported by:
    Social Science Planning Project of Chongqing, China (2025BS059)；General Program of the National Natural Science Foundation of China (52472339)。

PDF

摘要： 我国机动车保有量持续增长，城市道路交通突发事件呈现高频率和强扰动的特性，对道路交通运行监测和应急管理带来严峻挑战。传统交通事件监测体系在实时性、覆盖范围和信息维度上存在局限性，而社交媒体文本能够从公众感知角度提供细粒度事件信息。针对社交媒体交通突发事件文本噪声高、语义表达不规范、时空信息描述模糊等问题，本文提出一种融合深度语义建模与交通地理知识的交通突发事件提取方法。首先，搭建“Scrapy+Selenium”数据采集框架，制定交通突发事件文本采集规则，获取突发事件文本数据并完成数据预处理；其次，基于解耦注意力预训练模型（DeBERTa），实现对交通事件文本内容的准确识别；最终，构建融合地理知识的解耦注意力预训练模型-双向长短期记忆网络-条件随机场（DeBERTa-BiLSTM-CRF）命名实体识别模型，结合所建立的外部实体数据库，实现对交通突发事件时间与地点信息的精准提取。结果表明：DeBERTa分类模型F1值为0.9265；DeBERTa-BiLSTM-CRF模型在地点实体和时间实体的F1值分别为0.8595和0.8693，均优于对比模型。

关键词: 智能交通, 交通突发事件, 命名实体识别（NER）, 社交媒体, 深度学习, 文本分类

Abstract: With the continuous growth of motor vehicle ownership in China, urban road traffic incidents are exhibited the characteristics of high frequency and strong disturbance, which presents severe challenges to the operation monitoring and emergency management of road traffic. Traditional traffic incident monitoring systems have limitations in terms of real-time performance, spatial coverage, and information dimensionality, whereas social media texts offer fine-grained incident information from a public perception perspective. To address the issues such as high noise levels, non-standard semantic expressions, and vague spatiotemporal descriptions in social media texts on traffic incidents, this paper proposes a traffic incident extraction method that integrates deep semantic modeling and traffic geographic knowledge. First, a "Scrapy+Selenium" data collection framework is established, and acquisition rules for traffic incident texts are formulated to retrieve incident text data, followed by data preprocessing. Second, accurate identification of traffic incident text content is achieved based on the DeBERTa model with the disentangled attention mechanism and enhanced mask decoder. Finally, a geographic-knowledge-infused DeBERTa-BiLSTM-CRF named entity recognition (NER) model is constructed in combination with an established external entity database, which realizes the precise extraction of temporal and location information for traffic incidents. Experimental results demonstrate that the F1 score of the DeBERTa classification model reaches 0.926 5. For the DeBERTa-BiLSTM-CRF model, the F1 scores for locations and temporal entities are 0.859 5 and 0.869 3, respectively, both outperforming the baseline models.

Key words: intelligent transportation, traffic emergencies, named entity recognition (NER), social media, deep learning, text classification

中图分类号: