Road crashes and related fatalities are key development challenges in India. One of the critical barriers to understanding and investigating road crashes is the lack of good crash data. Most of the states in India do not have good road data; however, there are First Information Reports (FIRs) recorded for most road crashes, especially if they are fatal and grievous. The First Information Reports (FIRs) are crash descriptions recorded in text format by police in India. These FIRs are a potential source of crash data; however, only a few researchers have explored them because of their textual nature. This study bridges this research gap by applying natural language processing (NLP) tools (text mining, topic modeling, feature extraction) to FIRs data from Tamil Nadu state in India. The researchers have explored text mining on crash narration reports data in developed countries, but no study is available on FIR data of India. We did the feature extraction from FIR data using a linguistic rules-based approach. FIRs contains eleven essential features related to road crashes. These features are age, gender, detail of the victim's profession, date and time of the crash, causes of the crash, vehicle type involved, collision type, road facility type, body part injured during the crash, and IPC sections. This study uses 30890 FIRs from 2017 in Tamil Nadu for feature extraction. These extracted features are analyzed and compared with the Road Accident Data Management system (RADMS) to get a better understanding of road crashes and study the FIR data utility. These extracted features are also used to develop crash severity models with extracted features. We manually label 2299 FIRs into four crash severity types – fatal, grievous, minor, and no injury according to the standard definition of road crash severity. We apply two types of classification models – Random Forest and Multilayer Perceptron (MLP) for the road crash severity model. The Random Forest model performs well with a 91 percent F1 score.

To the best of our knowledge, it is the first study to reveal insights from the unstructured textual contents in FIR data. This study developed a framework of NLP tools that can be replicated in other languages of India for FIR data investigation. The feature extraction and road crash severity model can be extended to all states/languages of India, and significant insights can be obtained quickly with limited data. This study proves the usefulness of readily available FIRs data to understand the causes of severe road crashes and recommend mitigation measures. The findings of this study will contribute to a better understanding of the factors associated with road crashes.

KEYWORDS: Road crash, India, Road crash severity; First Information Reports; RADMS, NLP; Random Forest.