Zero-shot cross-lingual stance detection via adversarial language adaptation

Area of Science:

• Natural Language Processing
• Computational Social Science

Background:

• Stance detection typically focuses on single languages.
• Existing cross-lingual methods often require few-shot learning, limiting their applicability.
• Developing zero-shot cross-lingual models presents significant challenges.

Purpose of the Study:

• To introduce a novel approach for zero-shot cross-lingual stance detection.
• To enhance the performance of cross-lingual stance classifiers without target language training data.
• To address the limitations of current research in cross-lingual stance detection.

Main Methods:

• Proposed a novel approach: multilingual translation-augmented bidirectional encoder representations from Transformers (MTAB).
• Employed translation augmentation to improve zero-shot performance.
• Integrated adversarial learning to further boost model efficacy.

Main Results:

• Demonstrated the effectiveness of MTAB on vaccine stance detection datasets in English, German, French, and Italian.
• Showcased improved results compared to a strong baseline and ablated model versions.
• Validated the contributions of translation-augmented data and adversarial learning to model performance.

Conclusions:

• The proposed MTAB approach significantly improves zero-shot cross-lingual stance detection.
• Translation augmentation and adversarial learning are effective components for enhancing cross-lingual classification.
• The study provides a valuable contribution to multilingual NLP and social media analysis.