Syntactic Patterns in Political Speech: Dependency-Based Ad Hominem Detection in Bulgarian Social Media

The article by Nevena Grigorova, Stanislav Penkov, Keith Kiely, Dobromir Tsolyov, and Iglika Ivanova, titled “Syntactic Patterns in Political Speech: Dependency-Based Ad Hominem Detection in Bulgarian Social Media,” investigates how ad hominem attacks in Bulgarian political discourse can be automatically detected using syntactic structure, rather than relying exclusively on the words and expressions appearing in a text. The study is particularly relevant to Bulgarian as a low-resource and morphologically rich language, for which computational resources for fine-grained analysis of manipulative and argumentative discourse remain limited.

The study focuses on ad hominem as an argumentative strategy in which attention is directed toward a person’s character or credibility rather than the substance of their argument. The authors stress, however, that personal criticism is not automatically fallacious: criticism of a politician’s credentials, past conduct, or conflicts of interest may be relevant to an assessment of their reliability. This distinction makes automatic detection considerably more difficult than simply identifying insults or offensive vocabulary.

The empirical analysis uses a Bulgarian political social-media corpus collected through CrowdTangle and annotated for logical fallacies by six annotators. To improve label reliability, the study retains examples with at least 85% agreement among annotators. The resulting experimental dataset contains 177 sentences: 53 ad hominem examples and 124 control sentences, meaning that positive examples account for 29.9% of the dataset. All sentences are processed with the Bulgarian Universal Dependencies pipeline in Stanza.

A central contribution of the paper is the identification of five recurring syntactic configurations associated with ad hominem attacks. These include attributive insults, where a negative adjective or noun modifies the targeted person or group; predicative insults, where a negative characterization appears in the predicate; attacks embedded within complement clauses; coordinated discrediting, where several negative characteristics are accumulated; and passive discrediting constructions. The authors formalize these patterns through Universal Dependencies (UD) relations such as nsubj, amod, cop, ccomp, xcomp, conj, and nsubj:pass.

Importantly, the authors do not argue that these syntactic patterns automatically constitute ad hominem. The same grammatical structures can occur in neutral sentences. Their relevance emerges from the interaction between syntactic organization and negative evaluative content targeting a person, political party, institution, or social group. In this study, collective political entities are included because they can function as proxies for their members, representatives, or supporters.

For automatic classification, the study compares three types of representation: dependency-based structural features, lexical TF–IDF features based on word unigrams and bigrams, and a combined representation incorporating both. Three conventional machine-learning algorithms are evaluated—Logistic Regression, Decision Trees, and Support Vector Machines (SVM). Structural information includes counts and normalized ratios of UD relations, distributions of part-of-speech categories, and indicators approximating the five identified syntactic templates.

A particularly useful methodological point concerns evaluation on small datasets. An initial 80/20 train-test split makes structural features appear substantially stronger than TF–IDF in several configurations. However, when the authors apply their more robust primary evaluation—repeated stratified nested cross-validation with 5 folds × 20 repetitions—the difference becomes considerably smaller. This demonstrates how conclusions drawn from a single train-test split can be misleading when working with limited data.

Under this primary evaluation, TF–IDF with Logistic Regression provides the strongest overall baseline, achieving an average F1 of 0.548 ± 0.074 and ROC–AUC of 0.698 ± 0.093. Structural-only models remain competitive, with F1 scores of approximately 0.47–0.49, but they do not consistently outperform lexical features. Combining lexical and structural information also does not produce a consistent improvement in F1. Instead, structural information changes the precision–recall balance, suggesting that syntax contributes information complementary to lexical cues even when it does not increase overall classification performance.

Overall, the paper demonstrates that ad hominem attacks in Bulgarian political social media exhibit identifiable syntactic regularities that can support automatic detection. At the same time, the results caution against treating syntax as a replacement for lexical information. The more promising direction is to combine linguistic knowledge with lexical and contextual representations, particularly in low-resource settings where large annotated datasets are unavailable.

The authors also identify several important limitations. The dataset is small and imbalanced; classification operates at the sentence level and therefore lacks broader discourse context; the system does not perform coreference resolution; and informal social-media language can introduce errors into dependency parsing. The authors propose future work involving larger Bulgarian datasets, context-aware modelling, more sophisticated dependency paths and subgraphs, and the use of large language models as complementary tools for classification, annotation, validation, and structured explanation.

Suggested citation

Grigorova, Nevena, Stanislav Penkov, Keith Kiely, Dobromir Tsolyov, and Iglika Ivanova. 2026. “Syntactic Patterns in Political Speech: Dependency-Based Ad Hominem Detection in Bulgarian Social Media.” Proceedings of the Seventh International Conference on Computational Linguistics in Bulgaria (CLIB 2026), 386–395.

Article: Full-text PDF

Proceedings: CLIB Conference Proceedings