<![CDATA[Background: Social media has become one of the primary sources of information for the public, but it is also vulnerable to fake news (hoaxes). Various supervised learning approaches have been used for hoax detection; however, they generally depend on labeled data availability and require retraining to handle new claims. Objective: This study implements and evaluates a zero-shot inference-based hoax verification pipeline capable of verifying claims in near real-time without model retraining. Methods: The study uses a computational experiment approach consisting of four stages: (1) construction of a test claim dataset through paraphrasing using the NLLB-200, T5, and mT5 models; (2) keyword extraction using KeyBERT with IndoBERT embeddings and parameter optimization; (3) news article summarization using IndoBART; and (4) semantic similarity analysis by comparing Multilingual-E5, BGE, LaBSE, SBERT, IndoBERT, and TF-IDF as the baseline. Results: In the keyword extraction stage, the best balance between relevance, redundancy, and keyword coverage was achieved using the hybrid configuration of Top-k = 12, Top-p = 0.85, MMR = 0.6, and Ctx = 2. In the verification stage, Multilingual-E5 delivered the highest performance, achieving an accuracy of 0.96 and an F1-score of 0.94. In contrast, IndoBERT produced the lowest results, particularly for paraphrased claims. TF-IDF achieved good performance on original claims but experienced a decline when semantic variations were handled. Article summarization helped reduce article length, although this stage removed part of the information relevant to the verification process in some cases. Conclusion: The results of this study show that the zero-shot inference-, retrieval-, and semantic similarity-based approaches are capable of verifying hoaxes with high accuracy while maintaining stable performance across various claim variations. This approach offers greater flexibility for handling new claims than supervised learning methods and is suitable for application in near real-time hoax verification systems. Keywords: hoax verification, keyword extraction, semantic similarity, summarization, zero-shot inference]]>
Copyrights © 2026
