Sense and data scarcity: anticipatory mechanisms and sense completion in large language models / Смисл і нестача даних:механізми антиципації та смислової добудови у великих мовних моделях

The article presents a comprehensive analysis of the mechanisms by which large language models generate sense (sense completion) under conditions of data scarcity, fragmentation, or unevenness, drawing on the theory of human anticipation. The paper argues that while the processes of sense prediction in humans and the statistical data reconstruction procedures in neural network models differ in nature, they share several structural similarities: both systems aim to restore informational coherence, reduce uncertainty, and construct an internal model of context. Based on an examination of current research and original conceptual modeling, the study proposes a typology of six key sense-making mechanisms (sense completion) activated in data-limited conditions: a) probabilistic token prediction; b) implicit semantic interpolation; c) representational (supra-sense) generalization; d) hypercompensatory artifacts; e) contextual reconfiguration; and f) heuristic noise smoothing. The scientific novelty lies in formulating the first systematic typology of LLM sense-making mechanisms (sense completion) interpreted through an analytical framework grounded in human anticipation. It is demonstrated that these mechanisms can both enhance textual coherence and generate locally plausible yet globally invalid semantic decisions (statistically probable but semantically incorrect sense). Particular attention is given to explaining why the statistical nature of large language models makes them prone to “excessive anticipation” and erroneous reconstructions that lack counterparts in human cognitive strategies. The proposed classification allows for a clear distinction between constructive and risk-inducing types of sense-making (sense completion), identifies the preconditions for the emergence of false interpretations, and provides criteria for assessing the reliability of this process in incomplete or non-representative corpora. The findings form a theoretical basis for developing interpretable systems resilient to data scarcity and open a new direction in studying the anticipatory behavior of large language models within automated text analysis.