Mining Online Reviews with AI Sentiment Analysis: A Methods Guide for Agricultural Economists with Application to Affordable Wine Marketing

This paper demonstrates how agricultural economists can apply transformer-based sentiment analysis to online consumer reviews, providing both a methodological template and practical insights for wine marketing. We compile 11,415 consumer reviews of affordable wines (under $40) from major online retailers and apply a three-class RoBERTa transformer model to score sentiment at the sentence level. Theme-specific sentiment scores are extracted for eight wine attributes: value, acidity, sweetness, body, aroma, color, label/packaging, and taste descriptors. A multinomial logit model identifies which attribute perceptions most strongly predict positive versus negative or neutral reviews. Results show that 77% of reviews are positive, 12% neutral, and 11% negative—a distribution consistent with self-selection in purchase decisions yet containing sufficient variation for meaningful analysis. Marginal effects from the logit model reveal that for a typical negative reviewer, a 0.1-unit improvement in value sentiment increases the probability of a positive review by 7.7 percentage points, with taste (+6.9pp) and body (+6.0pp) showing similar effects. For wineries, the findings suggest that marketing affordable wines online should emphasize sensory attributes (acidity, body, aroma) rather than price, and that negative reviews typically require multi-attribute improvements rather than singledimension fixes. The methodology is reproducible with open-source tools and offers economists a cost-effective, scalable complement to traditional consumer panel surveys.