Student emotions, bad professors, and course ratings. A machine learning evaluation of one million student reviews.

For the first time in the literature, Natural Language Processing models based on deep neural networks are used to identify student emotions and extract bad performance by professors from the course reviews of students. I study how emotions and harmful performance are associated with high and low faculty ratings, the perceived level of difficulty of the course, and the university rating. I use a random sample of nearly one million student reviews from the ratemyprofessors.com website and the performance indicators of US universities from the Times Higher Education ranking. The linear and ordinal logistic regression models reveal strong relationships, and the estimated parameters have the expected signs. <br><br>Description of variables in the Results replication file, in Rdata format. 975,860 observations (rows). Raw data scraped from ratemyprofessors.com<br>- rev - textual student review<br>- rat - course rating (1-5)- dif - course perceived difficulty (1-5)- uname - University name- emo - emotion with the highest score- action_zs - professor bad performance with the highest probability calculated by zero-shot classification (DistilBERT model available at https://huggingface.co/typeform/distilbert-base-uncased-mnli)- sadness_sc - sadness emotion score calculated by DistilBERT model trained on Twitter corpus (available at https://huggingface.co/bhadresh-savani/distilbert-base-uncased-emotion)- similar for the other five basic emotions- nwords - number of words in the review<br> - m, y - month and year when the review was written<br><br><br><br><br><br>