Invented stories, distorted facts: fake news is spreading like wildfire on the internet and is often shared on without thought, particularly on social media. In response, Fraunhofer researchers have developed a system that automatically analyzes social media posts, deliberately filtering out fake news and disinformation. To do this, the tool analyzes both content and metadata, classifying it using machine learning techniques and drawing on user interaction to optimize the results as it goes.
Fake news is designed to provoke a specific response or incite agitation against an individual or a group of people. Its aim is to influence and manipulate public opinion on targeted topics of the day. This fake news can spread like wildfire over the internet, particularly on social media such as Facebook or Twitter. What is more, identifying it can be a tricky task. That is where a classification tool developed by the Fraunhofer Institute for Communication, Information Processing and Ergonomics FKIE comes in, automatically analyzing social media posts and processing vast quantities of data.
As well as processing text, the tool also factors metadata into its analysis and delivers its findings in visual form. “Our software focuses on Twitter and other websites. Tweets are where you find the links pointing to the web pages that contain the actual fake news. In other words, social media acts as a trigger, if you like. Fake news items are often hosted on websites designed to mimic the web presence of news agencies and can be difficult to distinguish from the genuine sites. In many cases, they will be based on official news items, but in which the wording has been altered,” explains Prof. Ulrich Schade of Fraunhofer FKIE, whose research group developed the tool.
Schade and his team begin the process by building libraries made up of serious news pieces and also texts that users have identified as fake news. These then form the learning sets used to train the system. To filter out fake news, the researchers employ machine learning techniques that automatically search for specific markers in texts and metadata. For instance, in a political context, it could be formulations or combinations of words that rarely occur in everyday language or in journalistic reporting, such as “the current chancellor of Germany.” Linguistic errors are also a red flag. This is particularly common when the author of the fake news was writing in a language other than their native tongue. In such cases, incorrect punctuation, spelling, verb forms or sentence structure are all warnings of a potential fake news item. Other indicators might include out-of-place expressions or cumbersome formulations.
“When we supply the system with an array of markers, the tool will teach itself to select the markers that work. Another decisive factor is choosing the machine learning approach that will deliver the best results. It’s a very time-consuming process, because you have to run the various algorithms with different combinations of markers,” says Schade.
Metadata yields vital clues
Metadata is also used as a marker. Indeed, it plays a crucial role in differentiating between authentic sources of information and fake news: For instance, how often are posts being issued, when is a tweet scheduled, and at what time? The timing of a post can be very telling. For instance, it can reveal the country and time zone of the originator of the news. A high send frequency suggests bots, which increases the probability of a fake news piece. Social bots send their links to a huge number of users, for instance to spread uncertainty among the public. An account’s connections and followers can also prove fertile ground for analysts.
This is because it allows researchers to build heat maps and graphs of send data, send frequency and follower networks. These network structures and their individual nodes can be used to calculate which node in the network circulated an item of fake news or initiated a fake news campaign.
Another feature of the automated tool is its ability to detect hate speech. Posts that pose as news but also include hate speech often link to fake news. “The important thing is to develop a marker capable of identifying clear cases of hate speech. Examples include expressions such as ‘political scum’ or ‘nigger’,” says the linguist and mathematician.
The researchers are able to adapt their system to various types of text in order to classify them. Both public bodies and businesses can use the tool to identify and combat fake news. “Our software can be personalized and trained to suit the needs of any customer. For public bodies, it can be a useful early warning system,” says Schade.