Researchers at the Universidad Europea del Atlántico (European Atlantic University, UNEATLANTICO), Dr. Mónica Gracia, international director of admissions at the Fundación Universitaria Iberoamericana (Iberoamerican University Foundation, FUNIBER), and Dr. Eduardo Silva, executive director of the same foundation in Guatemala, carry out a study with other professionals to improve automatic detection of fake news through machine learning. This work demonstrates the effectiveness of a hybrid and multivistic approach to analyze informational content in digital environments.
In recent years, disinformation has grown rapidly on social media and digital media. Messages that mix truth with falsehood influence public conversations, erode trust and complicate decision-making, especially in election periods or health crises such as COVID-19. Therefore, having automatic tools that help to identify and curb the dissemination of misleading content has become a priority for editorial staff, digital platforms and authorities.
Until now, many solutions have been based on examining only the text or using very complex models. Although they have made progress, they often find it difficult to adapt to new contexts, handle controversial texts or explain why they make a decision. They also tend to «flatten» all information into a single block, losing important nuances of language and subject matter.
The study proposes a different and easy-to-understand approach: to analyse each news item from three complementary angles and then combine these «opinions» in an intelligent way. First look at the text itself (frequent words and expressions), then how it is written (comprehensibility, emotional tone, use of proper names and grammatical structure) and finally what it really speaks about (dominant themes and general meaning). For each look a specialized model is trained; in the end, an «referee» brings all three together and decides more accurately than any individual.
To test it, the team worked with tens of thousands of items already classified as real or fake and used a rigorous evaluation that repeats training and testing in ten different rounds to avoid chance. He also checked whether the system remains stable when texts undergo minor alterations (for example, deleting or changing the order of some words) and whether it can transfer what has been learned to a different set of short political phrases.
The results are particularly strong. In the main set, the system hits 99.94% of the time and outperforms both models that look at a single perspective and others that mix everything in one step. It also improves to a very powerful reference based on deep neural networks. When evaluated with the set of short sentences, it maintains a very high level of accuracy (around 97%), indicating that it generalizes well even when the type of text changes. In stress tests, the accuracy remains above 97% even if some words are deleted, swapped or repeated, a sign that if one of the eyes loses information, the other two compensate.
This proposal strikes the right balance between effectiveness and computational cost: it improves key metrics without requiring heavy infrastructure, making it easier to adopt in environments with limited resources.
If you want to know more about this study, click here.
The Universidad Europea del Atlántico offers scholarships to study the Master of Data Science applied to Business Intelligence. This program is oriented to the design of analytical solutions that allow detecting patterns, evaluating informational risks and supporting evidence-based decisions.