Zero-shot Rumor Detection Across Languages And Domains

Social media platforms,as the primary medium for information dissemination,not only provide information to the public,but also facilitate the unprecedented spread of rumors,which affects people worldwide and poses a huge risk of misleading users and communities.Therefore,it is essential to quickly and accurately identify rumors,especially in the case of sudden topics and events that arise on global social media platforms.The cross-domain and cross-language zero-shot rumor detection task based on social media aims to predict the truthfulness of a given statement in scenarios where the topic domain of the training data and the test data is inconsistent and the language is different.Due to the significant uncertainty in the content of rumors,the response of social media communities with information propagation structures can provide clues for judging the truthfulness of statements.Recently,deep learning-based methods have shown better performance in rumor detection.This paper studies the zeroshot rumor detection task based on the propagation pattern structure and prompt learning framework of social media information,investigates and explores the modeling methods and encoding techniques for rumor detection tasks,and further analyzes the current research status and remaining problems of rumor detection tasks.The aim is to propose novel methods or architectures from both the modeling and encoding perspectives,which complement each other and promote the development of rumor detection tasks.The main contributions of this paper are as follows:1.Existing social media rumor detection methods either limit themselves to strict user response relationships or oversimplify social media conversation structures,leading to model performance being hindered by irrelevant interactions.In order to enhance user opinion interaction while mitigating the negative impact of irrelevant post interactions,and to capture the propagation chain posts in different topic domains,this paper proposes to encode the information propagation structure based on time and structural tree features.To increase the model’s awareness of global information and local subtree structures,we introduce absolute position encoding and relative position encoding based on propagation session thr-eads,and use an improved position-aware selfattention mechanism in the encoder module.Unlike traditional methods that integrate models with bottom-up and top-down modes,the proposed encoding structure allows for bidirectional interaction between posts with responsive parent-child or sibling relationships in the information propagation thread,thereby enhancing user opinion interaction and maintaining awareness of global structural information during rumor propagation.Experiments show that the rumor detection model based on social media propagation chain position encoding and position-aware selfattention networks performs well on three zero-shot rumor classification benchmark datasets.2.Although traditional pre-trained language model-based encoding techniques can learn representations of social media information,they still struggle to handle differences between different languages in cross-lingual scenarios.To address this issue,this paper proposes a rumor detection method based on unified prompt learning,which combines rumor detection and prompt learning and optimizes the model’s encoding for cross-lingual scenarios.Specifically,the lower layers of the pre-trained model are used to encode posts and prompt learning templates separately,capturing grammatical information that is highly relevant to their respective languages.These encoded representations are then jointly fed into the higher layers of the pre-trained model to model their abstract semantic information.Furthermore,the prompt words are optimized using contrastive learning to represent them in a latent space that is less dependent on language,thereby improving the effectiveness of crosslingual rumor detection.Extensive experiments on three zero-shot social media datasets demonstrate that this method further enhances the performance of rumor classification and exhibits excellent rumor detection capabilities in the early stages.