PROPer-BG: A New Publicly Accessible Package of Resources for Propaganda Detection in Bulgarian

The article by Irina Temnikova, Devora Kotseva, Ruslana Margova, Todor Kiriakov, Yolina Petrova, Nevena Grigorova, Alexander Komarov, Bogomil Katanov, Boryana Kostadinova, and Stefan Minkov, titled “PROPer-BG: A New Publicly Accessible Package of Resources for Propaganda Detection in Bulgarian,” introduces PROPer-BG, a publicly accessible package of datasets, taxonomies, and lexical resources designed to support the automatic detection and linguistic analysis of propaganda in Bulgarian. The work addresses a major gap in Bulgarian NLP: while propaganda-detection resources exist for several languages, Bulgarian remains comparatively under-resourced.

The package brings together two complementary resources: WASPer and PropaGATE. They differ in their sources, topics, annotation granularity, and balance between human-written and synthetically generated content. This combination is intended to support both machine-learning applications and linguistic research into the characteristics of propagandistic language in Bulgarian.

The WASPer dataset contains 1,696 short Bulgarian and English texts, combining human-written and AI-generated material. The Bulgarian training portion contains 734 examples, balanced between propaganda and non-propaganda, while a separate Bulgarian evaluation set contains 50 examples. The dataset distinguishes whether a text is human- or AI-generated, whether it contains propaganda, and, when propaganda is present, its category. Synthetic Bulgarian propaganda examples were generated using BgGPT and subsequently reviewed by domain experts.

An important part of WASPer is its four-level hierarchical taxonomy. It moves from identifying whether content is human- or AI-generated, through detecting the presence of propaganda, to assigning propaganda categories and specific techniques. At its most detailed level, the taxonomy contains 22 propaganda techniques grouped into five broader categories, including techniques such as appeal to fear, bandwagon, flag-waving, smears, stereotyping, appeal to authority, black-and-white thinking, whataboutism, and loaded language. The WASPer dataset had previously been used to train a propaganda-detection model achieving an F1 score of 0.853.

The second major component, PropaGATE, contains 1,939 synthetically generated, social-media-like Bulgarian texts manually annotated for whether they contain propaganda. A distinctive feature is that it deliberately includes contrasting viewpoints on the same topics. For example, propaganda can be generated both in favour of and against a particular ideological position. This design is intended to reduce the risk that machine-learning models simply learn that a particular topic or political position equals propaganda rather than learning the linguistic and rhetorical characteristics of propaganda itself.

This makes stance an important dimension of the resource. PropaGATE is presented as the first Bulgarian dataset of this kind to combine propaganda-presence annotations with partial propaganda-technique and stance annotations. In methodological terms, this is significant because it allows researchers to distinguish what position a text expresses from how that position is communicated—an important distinction when building systems intended to detect manipulative communication rather than particular opinions.

Another contribution is the creation of Bulgarian lexical resources for superlatives and intensifiers. The package includes a manually curated superlative lexicon containing 366 expressions—162 positive and 204 negative forms—as well as a second resource of 70 polarity-marked intensifiers. These resources are designed to support interpretable analysis of exaggerated, evaluative, and emotionally intensified language.

The linguistic analysis produces an important finding: intensifiers and superlatives are not reliable indicators of propaganda by themselves. What matters more is how they function in context. In WASPer, expressions equivalent to absolutely, terribly, completely, or incredibly can occur in both propagandistic and non-propagandistic texts. In propaganda, however, they more frequently amplify emotional or ideological evaluations, contribute to dramatization and delegitimization, or present opinions with exaggerated certainty. In non-propaganda texts, the same linguistic mechanisms tend to perform more descriptive functions.

The comparison between the two datasets also reveals different patterns of intensification. WASPer shows broader and more natural contextual use of intensifiers, whereas the synthetic PropaGATE data contains fewer such expressions but more explicit superlatives such as equivalents of the biggest or the worst. The authors therefore suggest that propaganda may depend less on obviously extreme vocabulary than on the interaction between lexical intensification, discourse context, loaded language, fear appeals, labelling, and other rhetorical strategies.

The authors are careful about the limitations and ethical interpretation of the resources. For human-written WASPer examples, the actual intention of the original authors cannot necessarily be established simply from the text. PropaGATE, meanwhile, is entirely synthetic and is not intended to represent actual people or events. The lexical analysis is exploratory rather than statistically validated, and intensification should be understood only as a proxy signal rather than definitive evidence of propaganda. The authors explicitly caution that PROPer-BG should not be treated as a tool capable of unambiguously determining whether a text is propaganda.

Overall, the paper’s main contribution is the creation of an open Bulgarian-language infrastructure for computational propaganda research. By combining human and synthetic data, stance-aware examples, a detailed propaganda taxonomy, and interpretable lexical resources, PROPer-BG can support the development and evaluation of Bulgarian propaganda-detection models while also enabling more detailed linguistic research into persuasive and manipulative communication.

Suggested citation

Temnikova, Irina, Devora Kotseva, Ruslana Margova, Todor Kiriakov, Yolina Petrova, Nevena Grigorova, Alexander Komarov, Bogomil Katanov, Boryana Kostadinova, and Stefan Minkov. 2026. “PROPer-BG: A New Publicly Accessible Package of Resources for Propaganda Detection in Bulgarian.” Proceedings of the Seventh International Conference on Computational Linguistics in Bulgaria (CLIB 2026), 14–27.

Article: Full-text PDF

PROPer-BG resources: Publicly accessible PROPer-BG package

Proceedings: CLIB Conference Proceedings