Sentiment Analysis of Teaching Evaluation Based on Complex Semantic Analysis

In the era of digital education, student evaluations of teachers constitute a crucial component of digital education. They serve as a driving force for promoting teaching reforms and are the fundamental basis for enhancing the quality of education and teaching. The digitalization of education reform provides students with more opportunities and avenues for evaluating teachers, including classroom teaching evaluations, online network teaching evaluations, and periodic assessments. Effectively utilizing this evaluation data to identify student needs and discover teaching issues is one of the effective ways to implement student-centered educational reforms. Through analysis, it has been found that existing student evaluation data exhibits characteristics such as implicit expression and complex emotional semantics, posing significant challenges for data analysis. This paper addresses these challenges by constructing a sentiment lexicon in the educational evaluation domain and employing complex semantic analysis to more accurately analyze the underlying emotional states within the evaluation data. The methodology involves the expansion of a general sentiment lexicon. Using active learning algorithms, sentiment seed words are selected from the evaluation data. Based on these seed words, an educational domain sentiment vocabulary is generated using the SO-PMI algorithm. The normalized educational domain sentiment vocabulary is then merged with the expanded general sentiment lexicon, resulting in the construction of the educational evaluation domain sentiment lexicon. During the evaluation phase, complex semantic analysis is applied to the evaluation data, and the educational evaluation domain sentiment lexicon is used for data analysis. Experimental results indicate that the proposed method achieves consistency with the actual data ranking in terms of sentiment classification and evaluation scores (MAE = 1.06, RMSE = 1.28). The F1 values for positive and negative teaching comments increased by 7.3% and 34.9%, respectively, compared to a general sentiment lexicon. Furthermore, when compared with common supervised learning algorithms, the proposed algorithm demonstrates superior sentiment classification performance.