Previous | 
Home
Section: New Results
Novel fact-checking architectures and algorithms
A frequent journalistic fact-checking scenario is concerned with the analysis of statements made by individuals, whether in public or in private contexts, and the propagation of information and hearsay (“who said/knew what when”), mostly in the public sphere (e.g., in discourses, statements to the media, or on public social networks such as Twitter), but also in private contexts (these become accessible to journalists through their sources). Inspired by our collaboration with fact-checking journalists from Le Monde, France's leading newspaper, we have described in [17] a Linked Data (RDF) model, endowed with formal foundations and semantics, for describing facts, statements, and beliefs. Our model combines temporal and belief dimensions to trace propagation of knowledge between agents along time, and can answer a large variety of interesting questions through RDF query evaluation. A preliminary feasibility study of our model incarnated in a corpus of tweets demonstrates its practical interest.
Based on the above model, we implemented and demonstrated BeLink [13], a prototype capable of storing such interconnected corpora, and answer powerful queries over them relying on SPARQL 1.1. The demo showcased the exploration of a rich real-data corpus built from Twitter and mainstream media, and interconnected through extraction of statements with their sources, time, and topics.
Statistic (numerical) data, e.g., on unemployment rates or immigrant populations, are hot fact-checking topics. In prior work, we have transformed a corpus of high-quality statistics from INSEE, the French national statistics institute, into an RDF dataset (Cao et al., Semantic Big Data Workshop, 2017, https://hal.inria.fr/hal-01583975), and shown how to locate inside the information most relevant to (thus, most likely to be useful to fact-check) a given keyword query (Cao et al., Web and Databases Workshop, 2018, https://hal.inria.fr/hal-01745768). Following on the above work, in [16], we present a novel approach to extract from text documents, e.g., online media articles, mentions of statistic entities from a reference source. A claim states that an entity has certain value, at a certain time. This completes a fact-checking pipeline from text, to the reference data closest to the claim. Using it, fact-checking journalists only have to interpret the difference between the claimed and the reference value. We evaluated our method on the INSEE reference dataset and show that it is efficient and effective. Further, this algorithm was adapted also to the (more challenging) context of content published on Twitter. This has lead to a semi-automatic interface for detecting statistic claims made in tweets and starting a semi-automatic fact-check of those claims, based on INSEE data. Figure 2 depicts the interface of this Twitter fact-checking system, which was shared with our Le Monde journalist partners.
