Towards automated real-time detection of misinformation on Twitter

Online Social Media (OSM) in general and more specifically micro-blogging site Twitter has outpaced the conventional news dissemination systems. It is often observed that news stories are first broken in Twitter space and then the electronic and print media take them up. However, the distributed structure and lack of moderation in Twitter compounded with the temptation of posting a news worthy story early on Twitter, makes the veracity of information (tweet) a major issue. Our work is an attempt to solve this problem by providing a approach to detect misinformation/rumors on Twitter in real-time automatically. We define a rumor as any information which is circulating in Twitter space and is not in agreement with the information from a credible source. For establishing credibility, our approach is based on the premise that verified News Channel accounts on Twitter would furnish more credible information as compared to the naive unverified account of user (public at large). Our approach has four key steps. Firstly, we extract live streaming tweets corresponding to Twitter trends, identify topics being talked about in each trend based on clustering using hashtags and then collect tweets for each topic. Secondly, we segregate the tweets for each topic based on whether its tweeter is a verified news channel or a general user. Thirdly, we calculate and compare the contextual and sentiment mismatch between tweets comprising of the same topic from verified Twitter accounts of News Channels and other unverified (general) users using semantic and sentiment analysis of the tweets. Lastly, we label the topic as a rumor based on the value of mismatch ratio, which reflects the degree of discrepancy between the news and public on that topic. Results show that a large amount of topics can be flagged as suspicious using this approach without involvement of any manual inspection. In order to validate our proposed algorithm, we implement a prototype called The Twitter Grapevine which targets rumor detection in the Indian domain. The prototype shows how a user can leverage this implementation to monitor the detected rumors using activity timeline, maps and tweet feed. User can also report the rumor as incorrect which can then be updated after manual inspection.