The practice of employing social media data as the monitor of public discourses and opinions has become increasingly popular with the development of big data techniques and machine learning tools. In this project, I intend to show that social media data can be used to refine the measurement of people’s gender-related attitudes, as a supplement to the traditional survey data. And more importantly, such finer-grinded and cheap-to-collect data may imply prevailing discriminations in local labor markets. To achieve this purpose, I trained a BERT classifier on a publicly available annotated misogyny dataset, and used it to select terms that are highly likely to be included in misogynistic tweets. In addition, I calculated weights for each terms as a measurement of how “efficient” they were in identifying misogynistic tweets. I then used the terms as well as their corresponding weights to construct a county-level Twitter Misogyny Index, and explored its implications on gender gaps in earnings and labor market participations. The results supported Becker (1957)’s taste-based discrimination model, and that each one standard deviation increase in the Twitter Misogyny Index is associated with a 0.02 increase in men-to-women median earnings ratio in case of all 16-year-old-and-above workers.