The comments teachers write when sending students to the office have the potential to increase our understanding of how bias may contribute to longstanding racial disparities in school discipline. However, large-scale analysis of open text has traditionally had a prohibitive cost. Through natural language processing techniques, we examined over 3.5 million office discipline records from national samples of more than 4,000 schools for whether teachers' linguistic patterns differed when describing incidents depending on the race/ethnicity and gender of the students. Results of such analyses consistently showed that teachers wrote longer descriptions and included more negative emotion when disciplining Black compared to White students, especially for Black girls. In conjunction with psychology of language theory, the patterns suggest that teachers may perceive and process student behavior differently depending on student identities. Implications of the findings and potential for research on naturally occurring language data in education are discussed.