Inspiration in human reasoning logic: Automating the inference and analysis of traffic accident information via...

Automated information parsing is crucial for analyzing the causes, responsibilities, and severity levels of traffic accidents from accident reports. The high costs of information annotation in existing studies and the limitations of traditional methods in handling texts with complex grammatical structures present significant challenges. Additionally, the cognitive reasoning and information transformation processes involved can be cumbersome. To address these issues, this study proposes a novel approach that enhances automation efficiency, ensures reasoning accuracy, and aligns with human logic in the automated analysis of traffic accident information.

Inspired by human reasoning logic (macroscopic grasp and microscopic depth), natural language processing (NLP) technology combined with highly comprehensible large language models (LLMs) has been utilized. Specifically, a dual-model collaborative system for automated traffic accident report processing is introduced by integrating topic identification through Latent Dirichlet allocation (LDA; Model 1 for a macroscopic analysis) and contextual understanding via fine-tuning bidirectional encoder representations from transformers (BERT; Model 2 for microscopic reasoning). The former is employed for accident causation analysis, while the latter is used for deducing accident severity. The system, LDA-Fine-tuning BERT Human Reasoning Synergy-of-Thoughts (LFBERT-ReasonSoT), is designed to extract information aligned with human reasoning logic.

In contrast to other classic models (Naïve Bayes, RUSBoost, Tree and Width Human-like Neural Network (WNN)), the LFBERT-ReasonSoT model exhibited superior performance. Specifically, it maintained an accuracy of 87.34% and achieved a positive predictive value of 87.81%, outperforming the Naïve Bayes, RUSBoost, Tree and WNN models across metrics, such as ACC, ER, TPR, TNR, PPV, FPR, F1, MCC, and KAP.

It aims to extract structured and inferable data from high-dimensional, large-sample, and unstructured complex accident text data, including temporal, geographical, positional, weather, and road condition features. It then identifies keywords, reasoning logic, and semantic relationships, enabling effective contextual analysis. This not only aids in determining accident causes and responsibilities but also accurately assesses accident severity. The findings of this study hold significant importance for the automated processing and conversion of text data, particularly in the context of highly complex and variable accident text data.