<p>Online training has become pivotal for teacher professional development, yet it faces challenges such as sentiment burnout and low learning achievement. To capture sentiment information non-invasively, this study employs a text-based sentiment analysis approach to examine the dynamic sentiment states of teacher-learners, thereby revealing the relationship between sentiment patterns and learning achievement and addressing a gap in the existing literature. To more accurately process large-scale text data, a semi-supervised machine learning method was applied to classify the sentiments of 7049 comments from 887 participants. Time series analysis was used to track the evolution of sentiments throughout the training process, while lag sequential analysis was employed to determine the sequential relationships and significant transitions between sentiments. The findings indicate that teacher-learner sentiments are predominantly neutral and positive. Over time, higher-achievement participants exhibited a clear shift from negative to positive sentiments, suggesting adaptive sentiment regulation and deeper cognitive engagement. In contrast, lower-achievement groups tended to maintain a neutral demeanor, potentially obscuring critical sentiment signals essential for timely instructional feedback. Based on these results, the study recommends that online teacher training programs enhance real-time monitoring and feedback of learners’ sentiments, encourage diversified sentiment expression, and optimize instructional interactions and course design through targeted sentiment interventions, ultimately improving training effectiveness.</p>