Incident Detection From Social Media Targeting Indian Traffic Scenario Using Transfer Learning

摘要

Road traffic congestion is one of the most challenging problems in densely populated cities. This paper aims to address this problem by developing a system to detect traffic congestion in India using Twitter. Twitter has been gaining momentum for research in congestion event detection for past several years because many commuters, as well as traffic authorities, tend to post traffic-related updates in real-time. There is no such traffic-tweet dataset for the Indian traffic scenario. We develop one such dataset that contains traffic-related posts concerning different Indian regions. The dataset contains posts that talk about traffic incidents such as accidents, infrastructure damage, and also about future planned events that can impact traffic flow. We call our dataset as L-TWITS (Labelled-TWeets for Indian Traffic Scenario). Basic practice in literature for traffic event detection problems is to collect a large amount of data, its annotation and then further analysis for event extraction. Such approaches often require a considerable amount of time for labelling the data. To address this shortcoming the proposed method uses a Transfer learning-based classifier that generally performs well even with less data. ULMFiT model has been used as a Transfer Learning approach for classifying the tweet samples into “Traffic incident related” or “Non-Traffic incident related” category. Experimental results on our labelled dataset show that ULMFiT outperforms other classification models making our model a convenient one for extracting traffic-related information targeting Indian scenario.