دانشگاه گیلان
معاونت پژوهش و فناوری دانشگاه گیلان
انتشارات دانشگاه گیلان

 آمار

تعداد نشریات33
تعداد شماره‌ها886
تعداد مقالات8,724
تعداد مشاهده مقاله55,279,201
تعداد دریافت فایل اصل مقاله10,515,594
Ziadid, Saeed, Hasheminezhad, Mahdieh, Mirroshandel, Seyed Abolghasem, Yousefi, Arash. (1405). LLM-based synthetic data augmentation for Persian sentiment classification. فیزیولوژی و بیوتکنولوژی آبزیان, (), -. doi: 10.22124/jmm.2026.32772.2992
Saeed Ziadid; Mahdieh Hasheminezhad; Seyed Abolghasem Mirroshandel; Arash Yousefi. "LLM-based synthetic data augmentation for Persian sentiment classification". فیزیولوژی و بیوتکنولوژی آبزیان, , , 1405, -. doi: 10.22124/jmm.2026.32772.2992
Ziadid, Saeed, Hasheminezhad, Mahdieh, Mirroshandel, Seyed Abolghasem, Yousefi, Arash. (1405). 'LLM-based synthetic data augmentation for Persian sentiment classification', فیزیولوژی و بیوتکنولوژی آبزیان, (), pp. -. doi: 10.22124/jmm.2026.32772.2992
Ziadid, Saeed, Hasheminezhad, Mahdieh, Mirroshandel, Seyed Abolghasem, Yousefi, Arash. LLM-based synthetic data augmentation for Persian sentiment classification. فیزیولوژی و بیوتکنولوژی آبزیان, 1405; (): -. doi: 10.22124/jmm.2026.32772.2992

LLM-based synthetic data augmentation for Persian sentiment classification

Journal of Mathematical Modeling
مقالات آماده انتشار، اصلاح شده برای چاپ، انتشار آنلاین از تاریخ 20 آبان 1405    اصل مقاله (1.18 M)
نوع مقاله: Research Article
شناسه دیجیتال (DOI): 10.22124/jmm.2026.32772.2992
نویسندگان
Saeed Ziadid1؛ Mahdieh Hasheminezhad* 2؛ Seyed Abolghasem Mirroshandel3؛ Arash Yousefi3
1Department of Computer Science, Faculty of Mathematical Sciences, Yazd University, Yazd, Iran
2Department of Computer Science, Faculty of Mathematical Sciences, Yazd University, Yazd, Iran
3Department of Computer Engineering, Faculty of Engineering, University of Guilan, Rasht, Iran
چکیده
Persian sentiment analysis faces significant challenges due to limited labeled data. Furthermore, the effectiveness of LLM-based synthetic data augmentation has not been systematically evaluated in this setting. We address this gap by leveraging DeepSeek-V3 and GPT-4o mini to generate synthetic samples across three distinct baseline configurations derived from SentiPers. Nine transformer architectures (Persian-specific and multilingual) were evaluated across three random seeds under a unified optimization framework. Rather than yielding uniform gains across all scenarios, the empirical utility of this technique proved highly task-dependent. While binary classification exhibited an apparent performance ceiling with marginal variations, five-class classification demonstrated substantial practical benefits, with synthetic expansion yielding observed performance gains in up to all nine model architectures. Rigorous empirical evaluation and intrinsic lexical audits confirm that targeted data injection enriches fine-grained lexical diversity, providing an effective low-resource strategy for resolving complex sentiment boundaries.
کلیدواژه‌ها
Persian sentiment analysis؛ language models؛ machine learning
آمار
تعداد مشاهده مقاله: 14
تعداد دریافت فایل اصل مقاله: 10


سامانه مدیریت نشریات علمی. قدرت گرفته از سیناوب