Road accidents are a worldwide problem, affecting millions of people annually. One way to reduce such accidents is to predict risk areas and alert drivers. Advanced research has been carried out on identifying accident-influencing factors and potential highway risk areas to mitigate the number of road accidents. Machine learning techniques have been used to build prediction models using a supervised classification based on a labeled dataset. In this work, we experimented with many machine learning algorithms to discover the best classifier for the Brazilian federal road hotspots associated with severe or nonsevere accident risk using several features. We tested with SVM, random forest, and a multi-layer perceptron neural network. The dataset contains a ten-year road accident report by the Brazilian Federal Highway Police. The feature set includes spatial footprint, weekday and time when the accident happened, road type, route, orientation, weather conditions, and accident type. The results were promising, and the neural network model provided the best results, achieving an accuracy of 83%, a precision of 84%, a recall of 83%, and an F1-score of 82%.
In this era of unprecedented digital connectivity and interactions, the issue of hate speech has become a focal point in societal discussions. The rise of digital communication platforms has fundamentally transformed how hate speech spreads. Online social media and messaging apps have rapidly disseminated hate speech, exacerbated by the internet's anonymity. Computational technology has emerged as a valuable tool for identifying and mitigating hate speech on social media. In this work, we employed five distinct corpora representing the English, Italian, Filipino, German, and Turkish languages. We propose employing a Large Language Model (GPT-3) enhanced with Cross-Lingual Learning to improve hate speech detection in English and Italian. Our investigation employs a strategy, namely JL/CL+, which combines two strategies: Joint Learning (JL) and Cascade Learning (CL). Even using data with lexical disparities, our findings demonstrate substantial success, yielding an F1-score of 96.58% for English and 92.05% for Italian languages.
29 September 2018 in Multimedia Tools and Applications