The frequency with which fake news or misinformation is published on social networks is constantly increasing. Users of social networks are confronted with many different posts every day, often with sensationalist titles and content of dubious veracity. The problem is particularly common in times of sensitive social or political situations, such as epidemics of contagious diseases or elections. As such messages can have an impact on democratic processes or cause panic among the population, many countries and the European Commission itself have recently stepped up their activities to combat disinformation campaigns on social networks. Since previous research has shown that there are no tools available to combat disinformation in the Croatian language, we proposed a framework to detect potentially misinforming content in the comments on social media. The case study was conducted with real public comments published on Croatian Facebook pages. The initial results of this framework were encouraging as it can successfully classify and detect disinformation content.

Detecting disinformation in morphologically rich and under-resourced languages remains a significant challenge in natural language processing. This paper examines the role of lemmatization as a preprocessing strategy for disinformation detection in Croatian. Building on previous work and extending the evaluation to a newly collected dataset of nearly 25,000 social media comments, we systematically compare traditional machine learning classifiers (SVM, Random Forest, and neural networks) and a transformer-based model (croBERT) on both original and lemmatized text. Our findings demonstrate that lemmatization does not produce uniform gains across architectures: while linear models and croBERT display small but measurable improvements from morphological normalization, non-linear models such as RBF SVM and neural networks experience substantial declines in performance. These results indicate that lemmatization interacts differently with model inductive biases and feature extraction mechanisms. Overall, the study provides a detailed empirical assessment of preprocessing choices for low-resource, morphologically complex languages and offers practical guidance for developing disinformation detection systems in Croatian and similar contexts.