Competitive Inference Memory Networks for Stance Detection and Evidence Extraction

Abstract

Due to the universality and the freedom of the Internet lots of fabricated messages are spread in various ways These messages often cause panic and chaos in the society Therefore fake news detection becomes a critical research field It verifies the news stories and rumors by observing different characteristics of false information Among various ways of detecting fake news one major way is to make use of the stance from the crowd to assess the credibility of the news Most of the methods take the stance as an important feature or utilize the stance changing over time to validate the rumors However how to precisely identify the stance of the news becomes the bottleneck of most models Thus the fake news stance detection is also regarded as a difficult issue At the same time previous researches only focused on predicting the stance of the news but ignored to provide favorable evidence Hence it is hard to convince users of the predicted result Therefore detecting stance with evidence become a growing research field in recent years In this paper we aim at detecting the stance of the news and extracting the descriptive evidence which can explain the stance simultaneously Recent systems are too dependent on the similarity between the news content and the claim to find favorable evidence The favorable evidence should contain more additional information to prove the stance about the claim Meanwhile they cannot take advantage of the evidence to infer the stance Thus we propose a method to find the words which have the same topics as the claim in the article by making a competition between words Moreover we utilize the interaction between these topic words and the claim to infer the stance Besides we use the topic words to assign a descriptive degree of each document toward the claim and extract the most descriptive document as evidence

Date of Award	2019
Original language	English
Supervisor	Hung-Yu Kao (Supervisor)