A Benchmark Dataset with Developing a Strong Baseline Accident Text Classification System for the Low-resource Bengali Language

The growing number of accidents in the world is a cause for concern. Accidents are one of the most tragic events in human life because people get injured, die, or suffer a lot. The rise of these incidents depends on various factors such as road conditions, human behavior, environmental changes, and more. Identifying and analyzing the accident causes and making awareness accordingly to the general people may be the principal tasks for minimizing and preventing accidents. We have found many prior works on identifying accident news and recognizing the accident causes from the text in English and other high-resource languages. However, researchers have little concern about classifying accident-cause from Bengali text data as a low-resource language. Therefore, firstly, we have developed a benchmark corpus named ACBT (Accident Classification from Bengali Text) of 3048 samples. Besides, we present a solid baseline accident detection (Sub-task A) and accident-cause identification (Sub-task B) systems by utilizing machine learning models. Our proposed voting ensemble method achieves the same highest F1-score and accuracy of 94.1% for Sub-task A, while for Sub-task B, an F1-score is 85.0%, and accuracy is 85.1%. Moreover, we believe our dataset might be beneficial in the reproducibility of the research to build an effective classification system. Thus, we publicly release the dataset at https://github.com/Nesarul-Hoque/Accident-Cause-Classification-from-Bengali-Text.