Automatic Classification of “Gross Negligence” in Serbian Court Decisions Using TF-IDF and Windowing

This paper presents an interpretable, resourceefficient baseline for detecting a concrete legal standard in Serbian judicial reasoning: whether a court decision affirms gross negligence or explicitly negates it. Using a small, expertlabeled corpus of decisions, we compare three input representations under leave-one-out evaluation: (i) manually extracted “gold” paragraphs, (ii) whole-document reasoning blocks, and (iii) window-based processing with document-level aggregation. Across experiments, the gold-paragraph setting achieves the highest performance, while the full-reasoning baseline outperforms the windowing approach on the current dataset. Windowing remains useful as an evidence-localizing representation that can support auditability and error analysis, even when it does not improve accuracy. The study is positioned as a pilot methodology and an engineering blueprint suitable for data-constrained e-Justice analytics.