Energy opinion classification models

The six models (one for each energy technology) are trained on a Tesla V100 GPU, by fine-tuning the RoBERTa transformer. The table below shows the size of the training set used for each technology, after balancing augmentation, and the model performance, evaluated using the Matthews Correlation Coefficient, ideal for binary classification problems.ModelNumber of training tweets(support ; oppose)Matthews Correlation CoefficientCoal(3918 ; 4074)0.74Natural Gas(2743 ; 2245)0.89Nuclear(9012 ; 10731)0.90Hydropower(7587 ; 5640)0.91Solar(13557 ; 13480)0.89Wind(15173 ; 15158)0.90The presented models may be improved upon, notably for coal energy classification. However, they can all be deemed to be quite efficient at correctly predicting the support category prevalence for a large-scale study as performed in this study. The model uncertainties are computed using the false positive and false negative rates observed during training, as indicated by the Matthew Correlation Coefficient measure.