Testing semantic settling dynamics model predictions – homonym meaning uncertainty

Abstract

Most of the semantic ambiguity research resulted in inconsistent findings. In order to account for those variations, Armstrong and Plaut (2016) suggested Semantic Settling Dynamics (SSD) model which predicts the dependence between ambiguity effects and time spent in semantic processing. Therefore, compared to the processing of unambiguous words, isolated polysemous words (multiple related senses) should show large facilitatory effect in early processing, which would decrease and eventually disappear. On the other hand, homonymous words (multiple unrelated meanings) should show no early effects, whereas inhibitory effect should appear later and increase during the course of processing. The model was tested behaviorally by comparing tasks with different processing demands. Typically, visual lexical decision task (LDT) was used to show effects expected in early processing, whereas some modification of this task was applied to prolong processing and reveal effects from later phases. Initial model testing in Serbian focused on polysemy described by a continuous measure – entropy (low entropy: a few senses, unbalanced sense probabilities; high entropy: many senses, balanced sense probabilities; Filipović Đurđević 2007, 2015). In this case, the model successfully predicted effect change in later processing (auditory LDT, compared to visual). The aim of this research was to test model prediction regarding homonymy by using the same design and same prolonging method. Furthermore, we tested two additional measures, components of entropy – number of meanings (NoM) and redundancy (low: balanced meaning probabilities; high: highly unbalanced probabilities). We presented 52 homonyms (2-5 meanings listed by native speakers) to 141 participants in two conditions, visual (baseline) and auditory (experimental) LDT. The results revealed that processing in the experimental condition (auditory LDT) was significantly longer as compared to the baseline condition (visual LDT). We observed the predicted interaction of task and redundancy: in visual LDT, there was no effect of redundancy (β = .021; S.E. = .012; df = 27.067; t = 1.661; p = .11), whereas facilitatory effect of redundancy was found in auditory LDT (β = -.052; S.E. = .019; df = 31.892; t = -2.785; p = .01). Consistent with previous results in Serbian (Filipović Đurđević, 2015) no entropy and no NoM effects were found. Crucially, the observed redundancy effects were completely in line with model prediction – no early effects (visual LDT), and facilitatory effect later in processing (auditory LDT). This facilitation stems from the fact that low redundancy homonyms are the polar opposite of unambiguous words. With an increase in redundancy, i.e. one of the meanings becoming more probable, a word will be processed more like an unambiguous word.