A FRAMEWORK FOR PROBABILISTIC RULE-BASED FAKE NEWS VERIFICATION

[[abstract]]社群媒體讓人們更容易獲取和分享信息。雖然信息可以通過報紙、電視或廣播獲得，但社群媒體使用起來更方便。然而，大量未經證實的消息在社群媒體上傳播，可能會誤導讀者，從而可能造成對抗，進而造成人與人之間的不信任或嚴重分裂我們的社會。它還可能擾亂國家經濟，降低人們對其政府的信心。 為了解決這個問題已經提出了許多方法。專業的事實核查員具有高精度驗證新聞的能力，以澄清信息是否為假。然而，極其耗時，無法滿足信息爆炸時代的巨大需求。因為專業的事實核查員的方法局限性而提出了對自動方法的需求。機器學習方法是一種可以對新聞驗證進行文本和語言特徵分類的方法。機器學習方法經常無法進行新聞驗證。這是因為所有新聞都沒有相同的寫作模式，導致無法運用機器學習所需的大數據分析技術。 本研究提出了一個框架，該框架採用基於概率規則的方法進行假新聞驗證。該框架非常靈活，可以通過輕鬆添加或刪除規則來擴展或修改。我們為此框架定義了一個自動化系統的過程。該過程由一系列的檢查點所組成。在這些檢查點當中會對新聞進行一一評價，得到最終評價結果，進而判斷該新聞為真的概率。我們利用PolitiFact 中新聞評論員對某些新聞的真偽評斷來測試我們的系統，結果顯示兩者結果是一致的。 Social media makes people lives easier to obtain and share information. Although the information can be obtained through newspapers, television, or radio, social media is more convenient to use. However, a lot of unverified news spread in social media may mislead readers, as a consequence, may create confrontation, which, distrust among people or severely divide our society. It disrupts the economy of countries, reduces people’s confidence in their government. Many approaches have been proposed to solve the problem. Professional fact-checkers have a capability with high accuracy to verify the news to clarify whether the information is fake or not. However, extremely time-consuming in manual verification may result in the difficulty to meet the huge demand in the era of information explosion. The limitation of the expert facts-checkers method raises the demand for an automatic approach. The machine learning approach has been applied to do textual and linguistic feature classification for news verification. However, the machine learning approach has not gained successful progress yet. It is because all news does not have the same writing pattern and structure that hinders it from applying big data analysis techniques required by the machine learning technology. This study proposes a framework that adopts a probabilistic rule-based methodology for fake news verification. This framework is flexible that can be expanded or modified by adding or removing rules easily. Our framework defines a process of an automatic system. The process consists of a series of checkpoints. A News is evaluated at these checkpoints, one by one, to get the final evaluation result which would tell the truth probability of given news. The proposed model is evaluated by comparing the obtained truth probability against the evaluation results performed by the journalist in Politifact.com. The comparison results show that our evaluation is consistent with the evaluation by Politifact.com