Their analyses show that semantic transparency is not a single simple trait. Measures that capture how much each modifier or head contributes to meaning behave differently from measures that capture how predictable the whole compound is. Those distinct dimensions also matter behaviorally: one of the factors, tied to the second constituent, reliably speeds up lexical decisions. That pattern gives a window into how readers assemble meanings in real time and how those processes vary across languages with different morphological structure.
For readers curious about language, learning, and inclusive design of language technologies, this work offers concrete implications. The dataset and methodological approach make it easier to combine human and computational perspectives when studying word meaning, and they suggest new ways to build models and educational tools that respect the multiple ways meaning is built. Follow the full article to see how these findings might reshape models of reading, lexical access, and tools that support multilingual users.
Abstract
Semantic transparency is a key construct for understanding how complex words are represented and processed, yet it has been conceptualized and operationalized in diverse ways across studies. In this study, we validate whether semantic transparency exhibits multidimensional properties across different measures in Mandarin Chinese. We first construct a novel dataset consisting of 2675 nominal compounds, with a rich set of measures from human ratings, traditional distributional semantic models, and recent large language models. To investigate whether they inform the same aspects of this construct, we then examine the latent structure among these measures through exploratory factor analysis. Our factor analysis reveals that this construct is fundamentally multidimensional, with measures assessing the semantic contribution of each constituent and the semantic predictability of overall compounds representing distinct factors in the latent structure. These derived composite factors also predict lexical decision performance, with the factor representing second constituent contribution showing significant facilitatory effects. Our work extends the cross-linguistic validity of the multidimensionality hypothesis of this theoretical construct previously established in English and German to Chinese compounds. Additionally, we provide a valuable resource for future research on the representation and processing of compounds, together with methodological insights into using computational estimates to augment psycholinguistic datasets across dimensions of semantic transparency.
