Accident event detection from Facebook posts written in Bengali and Banglish languages

Abstract

SummaryAccident events refer to abrupt or unpredictable incidents, which often cause a temporary or permanent bruise. Sometimes, the damage is so severe that it permanently ruins an individual's whole life and family. Detection of these unwanted accident events is the prerequisite for preventing and lowering the rate of casualty. Most people are habituated to posting their seeings, thoughts, and beliefs on social media and often post about accident incidents. This work detects accident incidents from real‐time Facebook data and the text‐mentioned probable time and location. Previous research works in this domain have been conducted in English, mainly with Twitter data. In this work, Bengali and Banglish Facebook posts have been used to aid the emergency response team in rapidly rescuing injured personnel. For this purpose, real‐time Facebook data have been crawled and pre‐processed. After that, several layers of feature extraction including individual keyword feature and paired keyword feature have been conducted and then sentiment analysis of the posts have taken place. Individual keyword feature is further divided into common event keyword and accident‐specific event keyword. Finally, the Support Vector Machine (SVM) classifier is used to classify accident and non‐accident events. Two more classifiers, Naive Bayes (NB) and Decision Tree (DT), are used for comparison. For NB, the Bernoulli, Gaussian, and Multinomial NB are applied. The SVM method achieves slightly better accuracy in the Bengali dataset than the Banglish dataset. The SVM, Bernoulli NB, Multinomial NB, Gaussian NB, and DT classifier achieved accuracy of 80%, 81%, 74%, 80%, and 79.5% for the Bengali dataset and 78%, 77.5%, 73%, 78%, and 78.5% for the Banglish dataset, respectively.