Semantic transparency in compound word reading: A computational investigation

Compound words (e.g., “catfish”) are combinations of constituent words that themselves possess free-standing meanings. Since the meaning of a compound is often related to the meaning of its constituents (i.e., the compound is semantically transparent; e.g., “snowball”), most compounds can be readily understood by a process of automatic composition. Indeed, evidence suggests that compound word processing involves both the attempt to retrieve its meaning from long-term memory (if the word is familiar) and to derive it from automatic constituent combination. In this context, many studies have relied on the behavioral effects of semantic transparency which, however, have been inconsistent, with some studies reporting facilitatory effects of transparency and others failing to detect any effect. To explain these inconsistencies, a recent study [1] showed that transparency effects may depend on participants reading experience. Concurrently, other studies suggest that transparency effects can be detected when defined “compositionally”, i.e., when transparency measures are obtained from computational models that model meaning composition explicitly. The present work brings together these perspectives by testing the interaction between compositional measures of semantic transparency and measures of reading experience. Measures of transparency were obtained from the CAOSS model [2], a computational model of compounding based on distributional semantics representations. This model learns to exploit the statistical regularities of compounding to generate a distributed representation of the predicted compound. Importantly, the CAOSS model allowed us to dissociate two factors underlying transparency: semantic composition and semantic consistency. Semantic composition captures the extent to which compound meaning can be predicted by constituent combination, while consistency captures the amount of meaning transformation required to turn the original constituent meaning into one that is relevant for compounding. We defined the former as the degree of match between model predictions and actual compound representations, and the latter as the magnitude of the transformation performed by the model. We employed these measures to predict eye movement data from CompLex [3], a large database of compound word reading. We find that semantic composition and consistency predict fixation times and interact with measures of reading experience. First, only experienced readers benefit from semantic consistency, while less experienced readers face an opposite, inhibitory effect of consistency. Second, we find that facilitatory effects of semantic composition emerge only for high-frequency compounds. These results show that semantic transparency effects on compound word reading arise from the complex interaction of individual- as well as item-level differences. We argue that computational modelling, by simultaneously disentangling variables of interest and providing data-driven measures, offers a compelling perspective to tackle and explain these complex phenomena.