Enhancing Vietnamese Stance Detection: Overcoming BERT Token Limits by Leveraging Text Summarization

This paper presents a work on detecting the stances of news readers towards a specific opinion/topic expressed in Vietnamese newspaper articles. We introduce a new architecture leveraging text summarization with a powerful transformer-based language model that excels in many downstream NLP tasks. A key feature of this architecture is its ability to handle transformer-based models' token limits by extracting the main ideas expressed by newspaper s', authors. To achieve this, we utilize the newspapers' titles/headlines as references to identify the important sentences within newspapers' bodies which results in maximizing the recall of ROUGE score. Experimental results on a Vietnamese benchmark dataset demonstrate the effectiveness of the proposed method. For the best method, we achieve a new state-of-the-art result by increasing the F1 score by approximately 2% in comparison to the method that does not use this simple yet effective text summarization technique.