Teaching Feedback Classification Protocol Benchmarked for Durability and Cross-Language Transfer

What changed

This research introduces a benchmark to assess the durability and cross-language transfer capabilities of a previously established protocol for classifying open-ended teaching evaluation feedback. The original protocol, validated on a Spanish institutional corpus using a frozen 2019-era transformer encoder, is re-evaluated across three distinct representation generations: sparse lexical features, frozen transformer embeddings, and prompted large language models (LLMs). The study also extends the protocol's sentiment classification task to English, using a balanced corpus of 45,000 comments that was cross-checked against an independently annotated aspect-based education dataset.

Key findings indicate that the protocol remains durable. A 2026 frontier model achieved the highest thematic F1 score on the most challenging Spanish task. However, this advanced model did not demonstrate a sentiment advantage over a more economical model, nor was there a significant descriptive separation between them on the English sentiment task. This suggests that the choice of model is primarily a deployment decision rather than an inherent property of the classification method itself.

The contributions of this work include:

• Durability Benchmark: The validated protocol was re-run on a Spanish teaching-feedback corpus using sparse lexical features, frozen transformer embeddings, and prompted LLMs. Cost, latency, and auditability were reported alongside F1 scores.
• Cross-Language Transfer Evaluation: The sentiment task was transferred to English, with a clearly defined scope and a documented labeling rule checked against an independent dataset.
• Reproducible English Sentiment Sample: A balanced English sentiment sample of 45,000 comments was created and released with its build recipe (script, seed, split indices) for exact reproducibility.

Why it matters for builders

This work provides a robust, validated methodology for processing and understanding large volumes of qualitative teaching feedback, a data source often underutilized by institutions. Builders can adopt or adapt this protocol to create more effective NLP tools for educational settings, ensuring that the analysis is reliable, auditable, and can extend across languages. The findings also guide model selection, emphasizing that the protocol's integrity is paramount, and advanced models may not always yield superior results for specific tasks like sentiment analysis.

Practical impact

Institutions can implement this protocol to systematically analyze student feedback, moving beyond manual reading or basic quantitative metrics. For developers, this means a proven framework for building classification systems that are not only accurate but also transparent and reproducible. The cross-language transfer capability is particularly valuable for global educational platforms or institutions with diverse student populations. Builders should consider integrating this protocol into their educational NLP pipelines, focusing on the annotation guide, reliability measurement, and validation design as core components, and then experimenting with different model architectures based on deployment constraints and specific task requirements.

Caveats and source limits

The research focuses on the durability and transferability of a specific classification protocol rather than introducing a new model. While modern LLMs were tested, the study highlights that they did not universally outperform older methods on sentiment tasks in English, suggesting that model choice is context-dependent. The benchmark was conducted on institutional teaching-evaluation feedback, and its performance on other types of text may vary. The paper does not provide specific pricing or detailed performance metrics for the tested LLMs, focusing instead on the protocol's robustness. The source is a research paper, and practical implementation details beyond the protocol's description are not elaborated.