Abstract: Education-domain sentiment lexicon (EDSL) was constructed based on multi-feature fusion to rectify insufficient domain-specific vocabulary coverage and contextual semantic shift associated with existing analysis tools commonly used for evaluation of massive open online courses (MOOC)． A word-expansion tool integrating cosine similarity algorithm and polarity orientation pointwise mutual information (POPMI) algorithm was developed, achieving expansion of potential domain-specific sentiment words from semantic similarity in vector space and statistical lexical co-occurrence association． Bidirectional encoder representations from transformer models were fine-tuned using domain-specific educational corpus, to obtain a sentiment polarity classifier with deep semantic perception capabilities, for automatic and accurate annotation of sentiment orientation of candidate words． General sentiment resources and domain-specific vocabulary were integrated through a fusion strategy, to form a highly adaptable educational sentiment lexicon． Compared to mainstream general sentiment lexicons, accuracy, precision, recall and F1 score these four indicators have shown significant improvement． EDSL provides effective affective computing tool support for intelligent teaching quality assessment and holds significant implications for advancing data-driven teaching improvements．