ViNSV is a system developed and designed to address the challenges of fact extraction and verification. This study presents the ViNSV system, which was developed as part of a competition UIT-DSC 2023 using the ISE-DSC01 dataset. The system leverages advanced natural language processing techniques, including BM25, Sentence-BERT, and XLM-R, to achieve high accuracy in fact extraction and verification tasks.
During a private evaluation conducted on the ISE-DSC01 dataset, the ViNSV system demonstrated an impressive Strict Accuracy of 76.33%. This outstanding performance earned the system a top 4 ranking, showcasing its remarkable competitiveness and effectiveness.
Figure 1: Classification Pipeline.
Figure 2: Evidence Retrieval Pipeline.
In addition, we have also demonstrated that the ensemble model provides higher performance. The ensemble model's performance was rigorously evaluated against the standalone BM25 and SBERT models, utilizing Top-1 Accuracy (Acc@1) as the primary metric for Support (S) and Refute (R) evidence categories (Fig. 3).
Figure 3: The performance of algorithms for the task of Evidence
Retrieval, with the Acc@1 metric for Support and Refute.
For any inquiries or issues regarding, please contact the following emails:
- Tran Quang Duy: 21522013@gm.uit.edu.vn
- Tran Thai Hoa: 21522082@gm.uit.edu.vn