Comparison of language models for wine sentiment analysis

Abstract This study presents a comparative evaluation of sentiment analysis models applied to a large corpus of expert wine reviews from Wine Spectator, with the goal of classifying reviews into binary sentiment categories based on expert ratings. We assess six models: logistic regression, XGBoost, LSTM, BERT, the interpretable Attention-based Multiple Instance Classification (AMIC) model, and the generative language model LLAMA 3.1, highlighting their differences in accuracy, interpretability, and computational efficiency. While LLAMA 3.1 achieves the highest accuracy, its marginal improvement over AMIC and BERT comes at a significantly higher computational cost. Notably, AMIC matches the performance of pretrained large language models while offering superior interpretability, making it particularly effective for domain-specific tasks such as wine sentiment analysis. Through qualitative analysis of sentiment-bearing words, we demonstrate AMIC’s ability to uncover nuanced, context-dependent language patterns unique to wine reviews. These findings challenge the assumption of generative models’ universal superiority and underscore the importance of aligning model selection with domain-specific requirements, especially in applications where transparency and linguistic nuance are critical.