What students need to know about fact checking and source verifying with AI

The notion of checking facts is a complex one, so we need first to make an effort to distinguish the main categories/steps in fact-checking and then try to identify how AI can help shed some light into all of them. In the class, we focus on fact checking and disinformation identification coming mostly or mainly in the form of text with a particular focus on the dissemination in social media.

One important first point is that AI gives us the tools (“AI models”) to properly extract quantifiable information from the textual or visual resources and use it to assign a value to every resource, claim or fact that we want to evaluate (“check”). This allows us to compare our fact with either facts that we know are true or other false facts in order to evaluate the truthfulness of the item in check. This shifts the mentality with which we need to approach the problem, as we move from a black and white world to a more grey area where we assess based on a score. This in many cases necessitates the use of a human to make the final decision; what in Ai we call a “human-in-the-loop” situation.

A second important point is to understand that all AI models and tools work with some error. There is no AI model that will give a hundred percent (100%) of correct/successful answers. This means that errors can accumulate and thus provide erroneous results at the end. We need to be well aware of that, since and due to the complexity of the problem we tend to use multiple AI models to evaluate different aspects during the fact-checking process.

So, in that respect the students need to understand that we need to work as follows:

1. Identify the fact
   This is a crucial first step and poses significant challenges since current NLP tools cannot identify all possible aspects of human communication including different expressions, negation, idioms, humor, etc. For that students are presented and invited to take a deeper understanding in tools of claim extraction. This is presented through different visual and textual examples.
2. Use different aspects (aka features in AI) to identify disinformation
   Different studies have shown that disinformation carries specific characteristics. We need to understand those characteristics and then use Ai tools to extract them. Those characteristics include aspects like the relation between the title and the content, the intensity of the expressions used, the use of images or other visual aids to enhance the message, etc.
   Students need to be aware that we now have the necessary AI models to quantify (“measure”) those different features and how we synthesize these measurements to come up with a final (semi)automated decision.
3. Compare facts with facts from trustworthy sources
   This effort is probably the most difficult one because we need to do different things and all of them are equally important for the final goal.
   (i) We need to evaluate the trustworthiness of our sources (so to have something to compare with).
   (ii) We need to extract the claims to compare from our resources and the trustworthy ones.
   (iii) We need to be able to compare the two correctly and make a decision on how close or how far they are and if this is enough to identify them accordingly.
   As all AI models work with some error, this process might accumulate problems from all different sources.

Mastering all the necessary AI models and tools is a rather demanding and complex task. Students need to understand how AI models work and what they can provide before they engage into actually using them in actual situations. Results though seem extremely promising so far.

REFERENCES

Atanas Atanasov, Gianmarco De Francisci Morales and Preslav Nakov: Understanding the Roles of Political Trolls in Social Media. CoNLL 2019 

Pepa Atanasova, Preslav Nakov, Georgi Karadzhov, Mitra Mohtarami, Giovanni Da San Martino: Overview of the CLEF-2019 CheckThat! Lab: Automatic Identification and Verification of Claims. Task 1: Check-Worthiness. CLEF (Working Notes) 2019 

William Ferreira, Andreas Vlachos: Emergent: a novel data-set for stance classification. HLT-NAACL 2016: 1163-1168

Benjamin D. Horne, Sibel Adali: This Just In: Fake News Packs a Lot in Title, Uses Simpler, Repetitive Content in Text Body, More Similar to Satire than Real News. CoRR abs/1703.09398 (2017)

Yixin Nie, Haonan Chen, Mohit Bansal: Combining Fact Extraction and Verification with Neural Semantic Matching Networks. AAAI 2019: 6859-6866

Dean Pomerleau and Delip Rao: The Fake News Challenge: Exploring how artificial intelligence technologies could be leveraged to combat fake news. 2017. http://www.fakenewschallenge.org/.

Martin Potthast, Johannes Kiesel, Kevin Reinartz, Janek Bevendorff, Benno Stein: A Stylometric Inquiry into Hyperpartisan and Fake News. ACL (1) 2018: 231-240

Hannah Rashkin, Eunsol Choi, Jin Yea Jang, Svitlana Volkova, Yejin Choi: Truth of Varying Shades: Analyzing Language in Fake News and Political Fact-Checking. EMNLP 2017: 2931-2937

James Thorne, Andreas Vlachos, Christos Christodoulopoulos, Arpit Mittal: FEVER: a Large-scale Dataset for Fact Extraction and VERification. NAACL-HLT 2018: 809-819

Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, Yejin Choi: Defending Against Neural Fake News. CoRR abs/1905.12616 (2019)