Usually, when seeing a fake news article, we don’t need to perform meticulous fact-checking to realize something is wrong — the story just doesn’t look right. But what features of text contribute to the impression? Can we teach a machine learning algorithm to make this type of decision? This blog post covers research on automatic style-based credibility assessments, the recent achievements and limitations.
Given how an explosive growth of the misinformation ecosystem has coincided with an equally abrupt renewal of interest in Artificial Intelligence (AI) in recent years, it is no wonder many see the AI as a solution to the problems we face. The most obvious use-case comes from web platforms that publish user-generated content, especially social media platforms. How can these platforms avoid becoming intermediaries in spreading deliberately false and harmful information?
The common approach is to rely on signalling that a certain post is doubtful through explicit warning messages, such as “Some or all of the content shared in this Tweet is disputed and may be misleading” labels. But these can be applied manually only in a limited number of high-profile cases — how can we judge the credibility of hundreds of millions of messages sent every day? The obvious way to streamline this type of decision seems to be Machine Learning (ML).
Before we dive into AI and ML technology, it’s worth taking a pause and thinking how do we, humans, tackle this challenge? After all, the term ‘fake news’ may be a recent phenomenon, but people have been deciding on information trustworthiness — ‘Should I believe this?’ — since they started using language. Often, we validate a message by comparing its content to what we already know and assessing its veracity, or at least plausibility. In the context of modern news this approach is known as fact-checking, since it relies on checking whether the events described indeed happened, and some progress is happening in the effort to perform this process automatically. But do we really always go through the toil of such verification?
Take a look at these two made-up headlines:
- Grey’s party finances under investigation following John Yellow’s revelations
- BREAKING: Yellow Just DESTROYED Grey’s Party… Here’s How!
The articles they advertise may describe the same event, but you wouldn’t find them equally believable, right? The overly dramatic vocabulary, the uppercase words, the exclamation ending… Clearly, apart from what the text says, we also take into account how the information is conveyed. The more sensational or emotional the language or structure of text, the less credible it seems. And rightly so: the investigation of fake news shows that they deliberately try to stir emotions, which influences the writing style.
The judgement of ‘style’ may sound like something reserved for humans, but quantitative approaches to the problem have been applied for a long time under the notion of ‘stylometry’. The techniques for measuring personal writing characteristics of famous authors go back to the 19th century and were later famously used to determine authorship of The Federalist Papers. More recently, the focus has moved towards using ML to gauge specific characteristics of an author based on their text, including gender, age, political affiliation or education. Now, when the vast amounts of informal and personal messages are shared through social media, even the measurement of personality becomes possible. Sometimes it may be hard to say what a classifier is using exactly to make its decisions, and whether we could still call it ‘style’, but the prediction accuracy speaks for itself.
Several approaches to use the techniques above in the domain of news credibility have been published already. In 2016, a team from Rensselaer Polytechnic Institute was able to differentiate 110 fake news items from the legitimate ones using classic stylometric features and measuring language complexity and psychological associations. A plethora of differences was found, including the low-credibility documents being shorter, using less technical vocabulary, simpler words, fewer quotes, more repetitions and pronouns for personal sentiment. It’s worth noting an important role of titles of fake news, which, despite using simple language, convey a lot of information and attract attention through a variety of means (e.g. uppercase words).
Others looked at how stylometry can help to predict party orientation of news outlets. It turns out it’s hard to guess if a text is right- or left-leaning by such means, but quite doable to differentiate mainstream from hyper-partisan writing, no matter the direction. More recently, a study published at AAAI conference extends such efforts to a corpus of 100,000 documents from 223 sources with expert-assigned credibility. The results show that some of the features are typical for particular news sources, which makes it harder to assess credibility of articles coming from newly emerging fake news websites. Nevertheless, over 80% accuracy is possible thanks to some of the credibility indicators present across sources.
For example, fake news articles are more likely to include words
- implying judgement, referring to moral values and goals,
- about power, respect and authority,
- with strong positive or negative sentiment
The credible documents, on the other hand, include plenty of
- words and syntactic patterns used to quote external sources,
- expressions of time and numbers.
The research is very much ongoing, but results so far show that human’s ability to judge text credibility based on writing style can be translated to ML models. But does it mean similar approaches could solve our problems with misinformation in near future by automatically filtering the content published? The answer is ‘no’ for several reasons. Firstly, most models trained recently use the most widely available low-credibility content, which are fake news stories made up for financial incentive. Sensational writing style is part of their business model, based on grabbing as much attention (and ad clicks) as possible. But we shouldn’t forget politically motivated misinformation, which focuses on influencing opinion. This goal may be better served through reliable-looking, yet still false, reporting. Whether style-based solutions can work in such challenging cases, remains to be seen. Secondly, even if we had a perfect AI fake news detector available, its implementation in practice requires some serious thought. It’s been argued that using algorithms to automatically filter user-generated content is equal to censorship and violates human rights. And the Facebook warning messages, mentioned at the beginning, have at best a ‘modest’ influence on users’ beliefs.
To sum up, the style-based credibility assessment models, or even any AI-based algorithm, can only be part of the solution to the problem of low-credibility content. Whatever the technical means we deploy, human effort will remain central in the battle against misinformation.