Conversation-based Reasoning for Explainable Fake News Detection via Weakly Supervised Evidence Detection

Fake news or misinformation circulated on media outlets could seriously hurt user experience and hinder the healthy development of Internet economy. Recent years witness the explosion of fake news, which has grew to be a daunting issue in human community. Automatic fake news detection and related research have attracted increasing attention from various disciplines. However, most fake news detection methods learn latent feature representations based on neural networks, which makes them black boxes to classify a piece of news without giving any justification.

This project aims to explore a principled and intelligent way for explainable fake news detection. Although interpreting machine learning models have been extensively studied, explainable fake news detection only recently received attention. Existing studies suffer from two drawbacks: (i) some systems naively select keywords or sentences as evidences to determine the veracity of news. But such evidences are not understandable for users with limited domain expertise; and (ii) the evidences for explaining fake news detection are generated via supervised learning and requires sizeable semantically annotation data, which is a daunting task especially for trending news. To tackle these fundamental challenges, in this project, we propose to develop a new explainable fake news detection framework that can effectively digest massive web data, mine the dependencies among them and automatically detect evidences for news veracity reasoning and interpretation. To handle the paucity of understandable explanations, the problem will be approached from two perspectives: (i) we enhance news storyline representation with conversation structures by mining both shallow and hidden dependencies among relevant contexts, which embed valuable signals about credibility and evidence. And (ii) text summarization techniques will be extended to detect evidences from the heterogeneous conversation utilizing graph transformer algorithm. To solve the model training issue, we will design a weak supervised learning framework capable of detect evidences from massive web data with minimal annotations, the basic idea of which is to employ the accessible coarse label (e.g., news uncertainty) to infer the evidentiality for each piece of corresponding storyline. Finally, on top of the detected evidences, we propose a new approach for reasoning news veracity and providing human-understandable explanations via sentence ordering and coherence modeling.

With our rich research experience, beside the theoretical and mathematical analysis, we plan to conduct interdisciplinary research and implement the framework on some realistic applications. We will collaborate with our industrial partners such as Master Concept Limited to provide quality news dissemination.