To make fact-checking faster and more effective, we researched and developed prototypes of fact-checking supporting tools built on top of state-of-the-art AI-based models.
Following the in-depth study of fact-checkers’ needs and requirements, we identified four steps of the fact-checking process where AI can help most. For each of these steps, we trained novel AI models and built the tools on top of them.
Since CEDMO covers fact-checking organizations producing fact-checks in three local languages (Czech, Slovak and Polish), multilinguality played an important role in the AI-based supporting tools.
Check-worthy Document Detector
The check-worthy Document Detector tool allows fact-checkers to identify which social media posts may be worth fact-checking more efficiently. A fact-checker is able to specify a list of sources to be monitored (currently Telegram channels are supported). The tool regularly gathers data from the provided sources and applies AI models to predict document’s checkworthiness. These predictions help a fact-checker to pre-screen potentially extensive lists of posts and decide which posts are worth manual in-depth investigation.
Factual Claim Extraction
Factual Claim Extraction tool aims to extract potential check-worthy claims – short textual statements. A fact checker is able to provide a whole document or its textual part potentially containing the check-worthy claims, and upon request, the check-worthy claims are highlighted.
Fact-check Finder
In the Fact-check Finder tool, a fact-checker can provide a piece of text, such as a social media post, and upon request, automatic detection of the previously fact-checked claims is performed. Matching is done against the extensive continuously updated database of fact-checking articles in multiple (primarily European) languages.
Fact-check Finder is available to a broad community of fact-checkers as a part of the WeVerify/InVid/vera.ai browser plugin (in the part Tools – Search – Fact Check Semantic Search). Do not hesitate to try it!
Fact Search
In the Fact Search tool, a fact-checker can provide a sentence containing a check-worthy claim. Upon request, documents that could help assess its veracity are recommended using automatic detection. The tool includes a claim verification module trained to classify the input claim into one of the three veracity categories (supported, refuted or not enough information).
To support media literacy acitivites we developed a CEDMO game. The game serves two purposes: educational and data collection. The educational purpose of the game is to familiarize players with a variety of misinformative content and its characteristics. The data collection focuses on acquisition of so-called “adversarial examples”, i.e. content that is very usable in strengthening AI based tools.
The goal of the game player is to modify a given tweet without changing its real meaning, yet causing an existing automatic misinformation detection model to fail in recognizing that the tweet contains misinformation.
