Sentence embedding to improve rumour detection performance model

Abstract

Recently, most individuals have preferred accessing the most recent news via social media platforms like Twitter as their primary source of information. Moreover, Twitter enables users to post and distribute tweets quickly and unsupervised. As a result, Twitter has become a popular platform for disseminating false information, such as rumours. These rumours were then propagated as accurate and influenced public opinion and decision-making. The issue will arise when a decision or policy with substantial consequences is made based on rumours. To avoid the negative impacts of rumours, several researchers have attempted to detect them automatically as early as feasible. Previous studies employed supervised learning methods to identify Twitter rumours and relied on feature extraction algorithms to extract tweet content and context elements. However, manually extracting features is time-consuming and labour-intensive. To encode each tweet's sentence as a vector based on its contextual meaning, we proposed utilising Bidirectional Encoder Representation of Transformer (BERT) as a sentence embedding. We then used these vectors to train some classifier models to detect rumours. Finally, we compared the performance of BERT-based models to feature engineering-based models. We discovered that the suggested BERT-based model improved all parameters by around 10% compared to the feature engineering-based classification model.