A study re-evaluates a validated protocol for classifying open-ended teaching-evaluation feedback by thematic category and sentiment. The research tests whether the protocol, originally built on 2019-era frozen embeddings, remains effective as representation methods advance and transfers to English.
The authors ran the protocol on an original Spanish institutional corpus across three representation generations: sparse lexical features, frozen transformer embeddings, and prompted large language models. They also transferred the sentiment task to a balanced 45,000-comment English corpus checked against an aspect-labeled education dataset.
Results show the protocol is durable; a 2026 frontier model posted the highest thematic F1 on the hardest Spanish task. However, it showed no sentiment advantage over cheaper models and no descriptive separation from them on English, indicating that model choice is a deployment decision rather than a property of the method.