Datasets are critical for emotion analysis in the machine learning field. This study aims to explore emotion analysis datasets and related benchmarks in online learning, since, currently, there are very few studies that explore the same. We have scientifically labeled the topic and nine-category emotion of 4715 comment texts in online learning platforms using the "three-person voting label method" based on the "sentence-level" and multi-category labeling dimensions with our self-developed system. After testing the consistency of the labeling results using the "Fleiss Kappa" method, we found that the consistency of the dataset was about 0.51, representing a moderate strength of agreement. Based on the dataset, the prediction accuracy of the Long-Short Term Memory (LSTM) method is about 0.68. This dataset provides a benchmark for the multi-category emotion dataset in the Chinese online learning field. It can provide a basis for the subsequent solution of emotion analysis, monitoring, and intervention in the education field. It can also provide a reference for constructing subsequent datasets in the education field.