Multi-Dilated Attention Framework for Tibetan Public Opinion Analysis Based on ERNIE

Understanding public opinion in low-resource languages such as Tibetan presents significant challenges due to limited annotated data and complex linguistic structures. In this study, we propose a novel Multi-Dilated Attention Framework for Tibetan public opinion analysis, leveraging the contextual representation power of the pre-trained ERNIE model as the backbone. Our architecture incorporates multi-dilated convolutional and self-attention blocks in the encoder to efficiently capture both local and global semantic features. Additionally, a multi-dilated cross-attention mechanism in the decoder enhances the alignment between contextual information and sentiment or intent indicators. The model is trained and evaluated on a manually collected and annotated Tibetan public opinion dataset. Experimental results demonstrate that our approach consistently outperforms traditional architectures across multiple evaluation metrics, highlighting its effectiveness in low-resource sentiment analysis tasks. This work contributes a scalable and linguistically informed solution for Tibetan NLP, and underscores the potential of multi-scale attention mechanisms in morphologically rich language settings.