Cross-lingual ABSA transfers aspect-level sentiment tagging to low-resource languages, but Korean’s morphology and subword fragmentation make alignment difficult. We propose a multi-level contrastive framework on top of an mBERT tagger, adding aspec...
Cross-lingual ABSA transfers aspect-level sentiment tagging to low-resource languages, but Korean’s morphology and subword fragmentation make alignment difficult. We propose a multi-level contrastive framework on top of an mBERT tagger, adding aspect-category token contrast and aspect-conditioned sentence contrast. We also test EMA-scaled and uncertainty-based weighting for multi-objective training. On Korean, French, Spanish, Dutch, and Russian, translation-based baselines outperform zero-shot transfer, and our aspect-aware contrasts yield the strongest gains (best overall with aspect-category token contrast). Adaptive weighting shows mixed results, suggesting noisy and sparse contrastive signals in cross-lingual settings.