Speaker
Fritz Günther
(University of Tübingen)
Description
While some compounds, such as shorebird, are semantically transparent, others, such as ladybird, are semantically opaque. Typically, this semantic transparency (ST) has been operationalized from a relatedness perspective, where ST depends on the relation between a compound meaning and the meanings of its constituents as independent units. However, such an approach falls short in capturing the compositional side of compounding, a morphological process whose main purpose is generating new meanings. We argue that compositionality, which describes how accurately an actual compound meaning can be predicted from its constituents and the way they are combined, is a crucial determining component of ST. In the work presented here, we demonstrate that compositionality plays an important role in compound processing. We use the CAOSS model, a compositional model for distributional semantics, to compute compositional distributed representations for compounds. Compositionality is then operationalized as the similarity between these compositional representations and the observed distributed representations for the same compounds, as induced from a large corpus of natural language. We further compute the similarity between the compound constituents and the compositional compound representations as measures of the constituent contribution to the compositional meaning. We test the impact of these variables on lexical-decision latencies and fixation-durations in reading. For both lexical-decision times and first-fixation durations, we find facilitated processing if the contributions of both constituents are high, indicating an early effect of constituent semantics. On the other hand, the effect of compositionality dissociates across tasks: in lexical-decision times, a facilitating effect is observed for low values of compound compositionality; conversely, gaze durations are shorter for high values of compound compositionality. These findings show that compositionality is crucial to the understanding of ST effects in compound processing. However, the associated variables can have different effects depending on task requirements.
Primary author
Fritz Günther
(University of Tübingen)
Co-author
Marco Marelli
(University of Milano-Bicocca)
