Emotions play an important role in e-learning environments, and may affect learning outcomes. Therefore, automatic emotion recognition of student’s emotions using machine learning algorithms has become an emerging research topic. To accomplish this goal, the first step is to collect a corpus of labeled emotions, and analyze the emotion types in the corpus. This study collects a text corpus of emotion sentences in mathematics learning. Each sentence is then annotated to provide analysis results such as the linguistic features, proportions, annotator agreements, and annotation accuracy for different emotions.