Persuasion-technique detection aims to identify rhetorical strategies used to influence how in- formation, people or events are perceived. Automatic detection of these techniques can sup- port the analysis of persuasive and potentially manipulative language in news and political com- munication. Most previous work has focused on high-resource languages, while resources for Georgian remain limited. This thesis investigates persuasion-technique detection in Georgian and introduces a new manually annotated dataset of 461 news articles containing 2,740 paragraphs. The annotation is based on the SemEval-2023 persuasion-technique taxonomy and includes 22 fine-grained techniques grouped into six broader categories. The experiments evaluated multilingual and Georgian-specific pretrained encoders, includ- ing mBERT, XLM-R and HPLT LTG-BERT, together with general-purpose large language models using zero-shot and few-shot prompting. For supervised classification, different combi- nations of multilingual data, manually annotated Georgian data, synthetic Georgian examples, translated data and oversampling were tested. Among the fine-tuned encoders, XLM-R-large achieved the best overall performance, and a Georgian-focused configuration using only manually annotated and augmented Georgian data performed on par with the strongest configuration that also included multilingual training data. The introduction of manually annotated Georgian data significantly improved fine-grained classification over training on the multilingual SemEval data alone. Additional Georgian-language training data, introduced through synthetic generation and translation, and combined with oversampling to address label imbalance, further improved performance in several configura- tions, though the gains were not always statistically significant. Few-shot Gemini and DeepSeek achieved results comparable to the fine-tuned encoders, with different strengths across individ- ual persuasion techniques. Overall, the results show that Georgian persuasion-technique classification can be approached effectively with both fine-tuned multilingual encoders and prompted large language models, despite the limited availability of Georgian training data. The experiments also show that ad- ditional Georgian-language training data can be particularly useful in this low-resource setting, although the effect depends on the augmentation method and the persuasion technique. The new corpus provides a resource for further research on persuasion analysis in Georgian, includ- ing rare techniques and span-level detection.
Persuasion-technique detection aims to identify rhetorical strategies used to influence how in- formation, people or events are perceived. Automatic detection of these techniques can sup- port the analysis of persuasive and potentially manipulative language in news and political com- munication. Most previous work has focused on high-resource languages, while resources for Georgian remain limited. This thesis investigates persuasion-technique detection in Georgian and introduces a new manually annotated dataset of 461 news articles containing 2,740 paragraphs. The annotation is based on the SemEval-2023 persuasion-technique taxonomy and includes 22 fine-grained techniques grouped into six broader categories. The experiments evaluated multilingual and Georgian-specific pretrained encoders, includ- ing mBERT, XLM-R and HPLT LTG-BERT, together with general-purpose large language models using zero-shot and few-shot prompting. For supervised classification, different combi- nations of multilingual data, manually annotated Georgian data, synthetic Georgian examples, translated data and oversampling were tested. Among the fine-tuned encoders, XLM-R-large achieved the best overall performance, and a Georgian-focused configuration using only manually annotated and augmented Georgian data performed on par with the strongest configuration that also included multilingual training data. The introduction of manually annotated Georgian data significantly improved fine-grained classification over training on the multilingual SemEval data alone. Additional Georgian-language training data, introduced through synthetic generation and translation, and combined with oversampling to address label imbalance, further improved performance in several configura- tions, though the gains were not always statistically significant. Few-shot Gemini and DeepSeek achieved results comparable to the fine-tuned encoders, with different strengths across individ- ual persuasion techniques. Overall, the results show that Georgian persuasion-technique classification can be approached effectively with both fine-tuned multilingual encoders and prompted large language models, despite the limited availability of Georgian training data. The experiments also show that ad- ditional Georgian-language training data can be particularly useful in this low-resource setting, although the effect depends on the augmentation method and the persuasion technique. The new corpus provides a resource for further research on persuasion analysis in Georgian, includ- ing rare techniques and span-level detection.
Persuasion Technique Detection in Georgian: Corpus Construction and LLM Evaluation in a Low-Resource Setting
MATIKASHVILI, TAMARI
2025/2026
Abstract
Persuasion-technique detection aims to identify rhetorical strategies used to influence how in- formation, people or events are perceived. Automatic detection of these techniques can sup- port the analysis of persuasive and potentially manipulative language in news and political com- munication. Most previous work has focused on high-resource languages, while resources for Georgian remain limited. This thesis investigates persuasion-technique detection in Georgian and introduces a new manually annotated dataset of 461 news articles containing 2,740 paragraphs. The annotation is based on the SemEval-2023 persuasion-technique taxonomy and includes 22 fine-grained techniques grouped into six broader categories. The experiments evaluated multilingual and Georgian-specific pretrained encoders, includ- ing mBERT, XLM-R and HPLT LTG-BERT, together with general-purpose large language models using zero-shot and few-shot prompting. For supervised classification, different combi- nations of multilingual data, manually annotated Georgian data, synthetic Georgian examples, translated data and oversampling were tested. Among the fine-tuned encoders, XLM-R-large achieved the best overall performance, and a Georgian-focused configuration using only manually annotated and augmented Georgian data performed on par with the strongest configuration that also included multilingual training data. The introduction of manually annotated Georgian data significantly improved fine-grained classification over training on the multilingual SemEval data alone. Additional Georgian-language training data, introduced through synthetic generation and translation, and combined with oversampling to address label imbalance, further improved performance in several configura- tions, though the gains were not always statistically significant. Few-shot Gemini and DeepSeek achieved results comparable to the fine-tuned encoders, with different strengths across individ- ual persuasion techniques. Overall, the results show that Georgian persuasion-technique classification can be approached effectively with both fine-tuned multilingual encoders and prompted large language models, despite the limited availability of Georgian training data. The experiments also show that ad- ditional Georgian-language training data can be particularly useful in this low-resource setting, although the effect depends on the augmentation method and the persuasion technique. The new corpus provides a resource for further research on persuasion analysis in Georgian, includ- ing rare techniques and span-level detection.| File | Dimensione | Formato | |
|---|---|---|---|
|
Thesis_PDFA1.pdf
accesso aperto
Dimensione
1.21 MB
Formato
Adobe PDF
|
1.21 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/115845