Polysemy in context: an experimental test of the sense entropy effect

Abstract

Entropy is a measure of sense uncertainty within polysemous words. The more senses a word has and the more balanced their probabilities are, the sense uncertainty is higher, i.e. the word is more ambiguous (Filipović Đurđević & Kostić, 2016). Previous research revealed the facilitatory effect of entropy on processing latency in visual lexical decision task (VLDT; Filipović Đurđević & Kostić, 2021; Mišić & Filipović Đurđević, 2021). This is frequently attributed to simultaneous activation of multiple related senses. In this paper we test the underlying assumption that if we were to restrict semantic activation to a single sense the advantage of widespread activation of related senses (i.e. the facilitatory effect of entropy) would disappear. One way to achieve this is by putting polysemous words in contexts constructed to evoke specific senses. Our prediction is that the entropy effect would disappear in sentences targeting specific senses, but prevail in ones that manage to maintain semantic ambiguity. A total of 86 participants took part in the self-paced reading task. Each of 102 polysemous words was presented in three sentences - to evoke the dominant, one of the subordinate senses, and in a neutral context that did not elicit any specific sense, mimicking the VLDT situation. Sentences were presented using the Latin square design between participants and words. Linear mixed effect regression analysis revealed that, compared to the neutral context, target words were processed more slowly in the context pertaining to the dominant sense (b = -.070, SE = .018, t(7327.3) = 3.817, p<.001), and even more slowly in the context which evoked the subordinate sense (b = -.111, SE = .019, t(7327.5) = 6.232, p<.001). This was in accordance with our prediction that the context would reduce the level of ambiguity and cancel out the facilitatory effect of the related senses. This was further corroborated in the absence of the entropy effect on the processing of the target word in the sense-evoking contexts. Furthermore, we observed an interaction which suggested the reversal of the entropy effect in subordinate sense context as compared to the dominant sense context (b = .043, SE = .019, t(7327.4) = 2.338, p = .019), which is in accordance with the models regarding word ambiguity processing in context (Armstrong & Plaut, 2016). However, the effect of entropy was also absent in the neutral context asking for further elaboration of our understanding of processing of polysemous words.