Abstract :
[en] This talk addresses the long-standing question of whether semantic change in the lexicon is, to some extent, predictable. Despite the central role attributed to related issues such as regularity in semantic change within historical linguistics (see, e.g., Sweetser, 1990; Traugott & Dasher, 2002), the predictability of semantic change remains largely unresolved. The main challenge lies in meeting the requirements for a robust predictive model of semantic change: (a) identifying reliable predictors; (b) securing access to large-scale synchronic resources with broad lexical and conceptual coverage, which are necessary for statistical modeling; and (c) assembling extensive diachronic datasets documenting historical semantic shifts across languages, which are indispensable for evaluating the model’s predictive power.
In this talk, our focus is on the PERCEPTION–COGNITION domain. While extensive descriptive work has documented many such pathways within this domain (Evans & Wilkins, 2000; Vanhove, 2008; Georgakopoulos et al., 2022; Norcliffe & Majid, 2024), the question of whether they can be systematically predicted from synchronic data alone has remained largely open. Most studies in the relevant literature, though not specifically concerned with the PERCEPTION–COGNITION domain, have focused narrowly on the directionality of figurative extensions, proposing variables that predict asymmetries in metaphorical and metonymic mappings (e.g., Xu et al., 2017; Winter & Srinivasan, 2022). However, their diachronic evidence, while valuable, often derives from word-formation patterns (e.g., List, 2023) rather than full colexifications. Moreover, this line of research has not yet fully exploited newly available digitized lexical resources such as Lexibank (List et al., 2022), CLICS (Rzymski et al., 2020), and NoRaRe (Tjuka et al., 2022), which enable large-scale, data-driven exploration of meaning relations across languages.
We then revisit the issue of predictability in semantic change by comparing two theoretical approaches that infer diachronic semantic extensions from synchronic coexpression patterns (on the term, see Haspelmath, 2023). The first approach relies directly on coexpression patterns, applying causal inference algorithms (Dellert, 2019) to derive diachronic trajectories from synchronic distributions (Dellert, 2016). The second approach investigates meta-parameters of connected concepts — such as concreteness, imageability, frequency, and age of acquisition — as potential predictors of directionality (Winter & Srinivasan, 2022; Carling et al., 2023; Baumann & Hartmann, 2026).
Both approaches are applied to data from the PERCEPTION–COGNITION domain (for recent work on this domain, see Georgakopoulos et al., 2022; Norcliffe & Majid, 2024). The first step involves constructing a colexification network based on CLICS³ data (Rzymski et al., 2020). For the first approach, we apply the PC* algorithm (Spirtes et al., 2000) directly to the coexpression data (Dellert, 2024) in order to infer possible diachronic trajectories between concepts. For the second approach, each concept is matched with corresponding metadata from NoRaRe, which provides values for concreteness, imageability, frequency, age of acquisition, and other variables across multiple languages. These meta-parameters serve as potential predictors of semantic directionality. For instance, as shown in Table 1 (see Appendix), HEAR exhibits higher concreteness and frequency scores, as well as a lower age-of-acquisition score, than OBEY, while imageability remains relatively similar across the two meanings. Following proposals that semantic change tends to proceed from more concrete, frequent, and earlier-acquired meanings to more abstract and later-acquired ones (see Xu et al., 2017; Winter & Srinivasan, 2022; Baumann & Hartmann, 2026), these values predict HEAR as the source meaning and OBEY as the target. This pathway is well attested in the literature and serves here as a calibration case, while the approach is ultimately intended to extend principled predictions to less-documented extensions within the PERCEPTION–COGNITION domain.
To evaluate the predictions of both approaches, we turn to diachronic evidence from well-documented historical languages, notably Greek and (Ancient) Egyptian. Additional data are drawn from DatSemShift 3.0 (Zalizniak, 2018). By systematically comparing these two modeling approaches and validating their predictions against diachronic evidence, this study aims to advance our understanding of which models and parameters can best predict semantic change in the lexicon.
