DiffCTRG: A Diffusion Model for City-Level Traffic Report Generation

Abstract

With the accumulation of massive traffic data in the era of intelligent transportation, automated city-level traffic report generation (CTRG) has emerged as an essential application. The CTRG aims to automatically generate textual summaries of road conditions from traffic data. It can serve as a new auxiliary decision-making tool for traffic management, which helps alleviate the workload of traffic managers and further improves the efficiency of traffic system management. However, similar to traditional image captioning tasks, CTRG faces the challenge of modeling the associations between different modalities of data. In this article, inspired by the impressive cross-modal data generation capability of diffusion models, we introduce the first baseline for the CTRG task, named diffusion-based CTRG (DiffCTRG). In DiffCTRG, we map textual descriptions to a latent space using a split bidirectional encoder representations from transformers (BERT), which aims to bridge the discrepancy between continuous diffusion processes and discrete textual data. In addition, a re-inference strategy is employed to enhance the correlation between different tokens according to the textual properties during the inference process. Extensive experiments on the Beijing text-traffic (BjTT) dataset demonstrate the effectiveness of DiffCTRG in generating high-quality traffic reports, establishing a baseline for this task.