Why do cultural traits sometimes change (rapidly)? Which cultural traits are more susceptible to change? The systematic empirical analysis of such questions requires reliable measures of cultural traits at the local level and over long periods. Since standard sources do not contain such information, I have collected and geo-referenced a data set of thousands of local U.S. newspapers and apply modern tools from machine learning and natural language processing to measure culture from newspaper text. The assumption underlying this approach is that language in local newspapers reflects the local culture. Validation exercises show that the text-based measures are strongly predictive of the contemporary survey measures. The key advantage of the method is that it allows me to track local cultures back into the 18th century at daily frequency.