Beyond Keywords: Modeling the SemanticComplexity of Deceptive Communication onInstant Messaging Platforms

Instant messaging platforms such as Telegram enable rapid information exchange butalso facilitate deceptive messaging at scale. In this study, we examine Telegram spamdetection through a hierarchy of models that vary in linguistic modeling capacity, frominterpretable lexical baselines (Logistic Regression, Random Forest, LightGBM) tosequential (GRU) and context-aware transformer representations (ALBERT). Usinga harmonized preprocessing and evaluation pipeline on 20,348 labeled messages, wecompare predictive performance across metrics (F1, ROC–AUC, PR–AUC, calibra-tion) and assess pairwise differences via McNemar’s test with multiple-comparisoncorrection. Across all metrics, ALBERT achieves the strongest performance and sub-stantially improves spam-class detection relative to lexical models. This performancegap is consistent with the presence of a subset of deceptive messages whose signals areless concentrated in surface keywords and more distributed across context. However,improved performance may also reflect differences in model capacity and inductivebias, benefits from large-scale pretraining, and stronger handling of sparse patternsvia contextual and subword representations. Accordingly, we interpret the proposed“complex tier” as an operational characterization of lexically subtle spam in this cor-pus, and we suggest that keyword-based moderation may be insufficient on its own tocapture the full spectrum of deceptive messaging observed here.