A study evaluates the durability and cross-language transferability of a validated protocol for classifying open-ended teaching-evaluation feedback by thematic category and sentiment.

  • The researchers re-ran the protocol on original Spanish data using three representation generations: sparse lexical features, frozen transformer embeddings, and prompted large language models.
  • They transferred the sentiment task to English using a balanced corpus of 45,000 comments checked against an aspect-labeled education dataset.
  • A 2026 frontier model achieved the highest thematic F1 on the hardest Spanish task but showed no sentiment advantage over cheaper models.

The authors conclude that model choice is a deployment decision rather than a property of the method, as the protocol remains effective across different representation technologies and languages.