Polarity Drift and Feature Entanglement in Cross-Domain Sentiment Analysis: Evidence from Amazon Product Reviews Case Study

Abstract

Unsupervised Domain Adaptation (UDA) for sentiment analysis is dominated by adversarial alignment methods. Unfortunately, it is increasingly criticized for inducing spurious correlations and feature distortion. Despite these theoretical objections, quantifiable empirical evidence demonstrating precisely how and why models fail at the semantic and geometric levels remains scarce. This paper presents a rigorous pilot study that quantifies adaptation barriers across three independent lenses for a single semantically distant domain pair (Books→Electronics). Lexical Evidence using Log-odds analysis identifies words with Polarity Drift Scores |PDS| > 1.1, and the max is 1.97. This experiment uncovered lexical evidence demonstrating a systematic vocabulary-level polarity reversal. Decision-Theoretic Evidence using a Books-trained BERT-base-uncased baseline achieves indomain F1 = 0.8975 but drops to 0.8705 cross-domain, with asymmetric positive-class degradation (ΔF1-POS = 3.3% > ΔF1-NEG = 2.1%) correlating with the drift lexicon. The Proxy A-distance dA = 1.855/2.0 indicates high domain separability for this domain pair. Meanwhile, the Geometric Evidence using inter-domain MMD (0.1475) is 4.10× larger than the intra-domain sentiment gap (0.0360), and within-target sentiment cluster separation is 45% smaller than within-source, providing strong geometric evidence of representation entanglement. Collectively, these findings demonstrate that for semantically distant domain pairs with high vocabulary divergence, forced marginal distribution alignment faces fundamental challenges. Conclusions are framed as strong but limited evidence from this single-pair pilot study; generalisability to other domain pairs requires further investigation.