Classificação de tweets relevantes em situação de desastre
Resumo
Resumo: Entre as diversas categorias de tarefas atualmente viáveis para o Processamento de Linguagem Natural, a classificação de texto é uma das de maior destaque pela sua versatilidade entre domínios e aplicações especificas. Na consolidação global da utilização das redes sociais, nas quais inúmeras pessoas compartilham informações das mais diversas naturezas, encontra-se a aplicação de distinguir quais fatos são reais. No meio desse escopo, a classificação efetiva de noticias relativas a desastres e incidentes possui particular elevância, de forma a governos e órgãos de auxilio atuarem com a maior agilidade possível com o objetivo de reduzir e amenizar danos. Neste artigo são comparados o desempenho de 3 algoritmos clássicos de machine learning e um de deep learning (BERT), que possui representações contextuais treinadas em redes neurais profundas, na classificação de tweets relativos a desastres reais. Os resultados apresentados por meio de métricas de desempenho demonstram que o BERT supera significativamente os demais algoritmos nesta tarefa. Abstract: Among the high diversity of categories related to practicable tasks nowadays in Natural Language Processing, text classification is one of the highlights by the versatility amid domains and specific applications. By the global usage consolidation of social networks, where innumerable people share information about all sorts of nature, we find the application of determining which facts are real. In this scope, the effective classification
of news related to disasters and incidents has a particular relevance, in order to governments and rescue agencies act as soon as possible to reduce damages. In this article the performance of 3 classical machine learning algorithms and one deep learning (BERT), with contextual representations trained on deep neural networks, are compared for the classification of tweets related to real disasters. The results by performance metrics shows that BERT overcomes significantly the remaining algorithms in this task
Collections
- Data Science & Big Data [107]