A Dual-Annotated YouTube Comment Corpus for Studying Sentiment Classification Divergence in Emerging Technology Discourse

Abstract

This data report documents a corpus of 16,652 YouTube comments on precision fermentation technology, each independently classified for sentiment by two automated approaches: GPT-4o, a large language model accessed via prompt engineering, and EmoRoBERTa, a fine-tuned transformer model. The two models disagree on 39.1% of all classifications, and the disagreement is systematically asymmetric: EmoRoBERTa classifies substantially fewer comments as negative than GPT-4o. A subset of 301 comments includes additional human-annotated labels, providing indicative validation of model performance. The corpus spans February 2019 to November 2024, covering the period during which precision fermentation products entered commercial markets. To comply with YouTube API data redistribution policies, the deposited files contain comment identifiers and model-generated labels without comment text, following established rehydration conventions. The full data collection and preprocessing pipeline is documented in a public code repository. This corpus supports research on cross-model annotation reliability, sentiment classification methodology, and computational approaches to studying public discourse on emerging technologies.