Evaluative Framing of Mathematics in Serbian-Language Discourse: Topic Modeling and Sentiment Analysis Through a Control-Value Theory Lens

Mathematical education is shaped not only by classroom instruction but also by the broader evaluative narratives that circulate in everyday social discourse. Such narratives may inform perceptions of controllability and subjective task value in mathematics, and thereby bear on achievement emotions, academic engagement and persistence. In this study we examine how mathematics is evaluated in a Serbian-language web corpus, combining Control-Value Theory (CVT) with topic modeling and computational sentiment analysis. Topic modeling showed that discourse about mathematics is thematically diverse and centred on educational contexts, while also extending to economics, politics and other public domains. Sentiment was classified using a BERTić model fine-tuned on raw, unlemmatised text. To address class imbalance, minority classes were oversampled using a square-root-inverse weighting scheme, and model checkpoints were selected based on macro-F1, with logit adjustment applied at inference. In three-fold stratified cross-validation, this procedure yielded a macro-F1 of 0.65 (SD = 0.03), compared with 0.32 under the original single-split procedure. Given the sensitivity of performance to the modelling and evaluation procedure, sentiment-related findings are treated as exploratory rather than confirmatory. Applied to the 9000 unannotated segments, the classifier assigned 92.1% to the neutral class, 4.1% to the positive class and 3.9% to the negative class. A text-internal register analysis showed that explicitly evaluative segments become steadily more frequent as texts become less institutional, rising from 3.1% through 7.1% to 14.1% across three register strata, χ2(4) = 273.71, p < 0.001, Cramér’s V = 0.12. Because three-class polarity is a coarse indicator of the appraisal dimensions CVT specifies, the corpus was additionally coded with a dictionary of eight appraisal-relevant cue classes (difficulty, competence, controllability, success and failure, usefulness, relevance, high-stakes assessment and prestige). Appraisal-relevant framing proved far more common than explicit evaluative marking: 39.7% of segments carried at least one appraisal cue, against 7.8% carrying an explicit polarity label. Topic-conditioned polarity could be estimated only in part: of the six topics large enough to support an estimate, only one pair of bootstrap confidence intervals separates, so this analytic goal is reported as largely unmet. Predominant neutrality is therefore interpreted as a joint product of the register composition of srWaC and the ceiling of three-class polarity classification, and not as evidence that mathematics is affectively neutral in Serbian discourse. The study describes public representations of mathematics. It observes no learners, teachers or classrooms, and its implications for instructional design and teacher communication are stated as theoretically motivated hypotheses for future research rather than as empirically validated intervention targets.