Urban Transport Evaluation Using Knowledge Extracted from Social Media

• Autor(a) principal: Francisco André Barreiros Murçós
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: https://hdl.handle.net/10216/137319
• topic: Engenharia electrotécnica, electrónica e informática; Electrical engineering, Electronic engineering, Information engineering

Public opinion is nowadays a valuable data source for many sectors. Regarding the transportation and mobility sector, it is possible to collect information on real-time with reduced costs compared to other methods of information extraction. In this dissertation, we defined a methodology to extract knowledge from messages collected from Twitter to analyse urban mobility. The methodology was structured according three main modules: system configuration, data analytics and visualization. The messages used for the demonstration of the proposed methodology were extracted during two months for three different cities: New York, London and Melbourne. The text extraction from social media and its analysis are very time-consuming tasks due to the volume of the messages produced. Each message extracted from Twitter is, normally, short, informal and with a lot of slang or misspellings. To deal with that matter, by using NLTK (Natural Language Toolkit) tool, NLP (Natural Language Processing) techniques were applied so the text could be cleared and understandable by the algorithm. For the classification of travel related messages, a BERT (Bidirectional Transformers for Language Understanding) embedding model was used. The model is pre-trained, unsupervised and was released in 2018. In order to understand if a simple model could have good performance, an unigram approach was used. Three lists of travel-related words were used: (i) a small list with 10 traveled-related words, (ii) a medium list with 35 traveled-related words and (iii) a big list with 344 traveled-related words. The results show a high model performance with precision and accuracy higher than 0.80 and 0.90, respectively. Popular words are train, walk, street, car, station, street and avenue. Consistent results were obtained for all the three cities assessed. To evaluate the public opinion, the messages related to transportation and mobility were classified according to its sentiment. Then, to evaluate the polarity of the messages (positive, neutral or negative), VADER (Valence Aware Dictionary and sEntiment Reasone) sentiment tool was used. VADER is an easy tool to use and has great compatibility with social media messages and informal texts. It is a lexicon and rule based tool that calculates the compound value of text emotion according to its words. The developed methodology attained good performance results for the sentiment analysis where the average value of precision scored 0.77 while recall, accuracy and F1-score attained around 0.78. A specific analysis was made regarding a car crash event on New York on May 18, 2017. This analysis demonstrates that the methodology is capable of recognizing spacial changes and mobility flows directing to the potential causes of its origin. The developed work allows the conclusion that the proposed methodology can be very helpful to transport engineers, urban planners, researchers and policymakers in getting insight into public opinions regarding urban mobility.