Towards Reproducible Cross-Lingual Sentiment Transfer: A Systematic Framework for Persian English Using ParsBERT and XLM-RoBERTa

Cross-lingual sentiment transfer from English to Persian (and vice versa) remains challenging due to script differences, morphological richness, and limited parallel resources. This paper introduces a comprehensive, fully reproducible experimental framework for systematic comparison of three modeling paradigms: (۱) an enhanced rule-based lexicon baseline with negation and intensifier handling, (۲) classical TF-IDF + Logistic Regression and SVM, and (۳) fine-tuned transformer models (ParsBERT and XLM-RoBERTa). We define five controlled scenarios—monolingual (FA→FA, EN→EN), cross-lingual (EN→FA, FA→EN), and zero-shot (fine-tune on English, direct test on Persian)—using cleaned and class-balanced versions of SentiPers (۲۶,۷۶۷ sentences) and SemEval-۲۰۱۷ Task ۴ (English Twitter). The framework specifies scientific preprocessing (URL removal, Persian normalization, deduplication), stratified ۷۰/۱۵/۱۵ splits with fixed seeds, hyperparameter ranges, macro-averaged metrics + per-class F۱, McNemar’s statistical tests, structured error categorization (negation, sarcasm, mixed polarity, domain/OOV), and ablation studies on fine-tuning, normalization, and balancing. All code, splits, and seeds will be released publicly upon completion of the empirical phase. This blueprint addresses the reproducibility gap in low-resource cross-lingual NLP and provides a transparent roadmap for Persian–English sentiment transfer research.

کلیدواژه ها:

Cross-Lingual Sentiment Analysis ، Persian-English Transfer ، ParsBERT ، XLM-RoBERTa ، Reproducible Framework ، Ablation Study ، Error Analysis