Using an auxiliary dataset to improve emotion estimation in users’ opinions