What Makes Cross-Domain Sentiment Transfer Hard? Target Difficulty versus Transfer Distance in Nepal Tourism Reviews

Cross-domain sentiment transfer is usually framed as a distance problem: performance is expected to degrade with source–target dissimilarity. We test this on a new corpus of English-language Nepal tourism reviews with a 2×2 factorial structure (two cities × two review domains), evaluating all 16 source–target settings (four in-domain and twelve off-diagonal, or cross-cell, transfer settings) across six models. In a descriptive variance decomposition of the complete RoBERTa-base and DeBERTa-v3-base transfer matrices, target-cell identity accounts for 54.5% of the total sum of squares, while source-cell identity accounts for under 1%. The associated ANOVA p-values are nominal and interpreted cautiously because transfer-matrix entries share target test sets and are not independent replicates. Across the four evaluated targets, higher target macro-F1 (i.e., lower target difficulty) is associated descriptively with longer reviews and a larger share of negative-class reviews under macro-averaged evaluation; however, each per-cell test estimate for a given seed is based on one held-out place, so these associations may reflect venue-specific characteristics as well as cell-level differences. A separate shift-type analysis finds that domain shift is more damaging on average than geographic shift in this corpus. A separate pooled-training comparison finds no DANN improvement over the pooled RoBERTa baseline, but because both systems used training data from all four target cells, this is not evidence about transfer to an unseen target. These observational patterns do not establish a causal mechanism or population-level effects across destinations.