The persistent rise in road traffic fatalities continues to pose a grave public health and socioeconomic challenge globally. Accident documentation—while rich in contextual information—exists predominantly in free-form textual formats that resist scalable automated processing. Contemporary advances in Artificial Intelligence (AI), Natural Language Processing (NLP), and deep learning now afford promising pathways for converting such unstructured narratives into actionable safety intelligence. This survey investigates how Multi-Task Learning (MTL) architectures can be applied to concurrently address two intertwined objectives: identifying multiple co-occurring traffic violations (e.g., overspeeding, signal non-compliance, helmet absence, wrong-lane usage, and unauthorised parking) and estimating the severity of the resulting collisions. Looking across the studies we reviewed, no single model family dominates – transformer encoders handle text well, graph networks handle road topology well, and probabilistic models handle uncertainty well, but combining all three remains an open engineering challenge. Beyond raw prediction accuracy, the harder problem turns out to be making these systems trustworthy enough for real enforcement use- explainability tools out to be making these systems trustworthy enough for real enforcement use – explainability tools help, but they are not yet adequate for legal contexts. Beyond raw prediction accuracy, the harder problem turns out to be making these systems trustworthy enough for real enforcement use – explainability tools help, but they are not adequate for legal contexts. We close by identifying what the field most urgently needs: shared benchmark datasets, standardised evaluation metrics, and honest engagement with the fairness and accountability questions that deployment will inevitably raise.
V et al. (Sun,) studied this question.