A closer look at scalar diversity using contextualized semantic similarity

Abstract

We take a closer look at van Tiel et al.’s (2016) experimental results on diversity in scalar inference rates. In contrast to their finding that semantic similarity had no significant effect on scalar inference rates, we show that a sufficiently fine-grained notion of semantic similarity does have an effect: the more similar the two terms on a scale, the lower the scalar inference rate. Moreover, we show that a context-sensitive notion of semantic similarity (in particular ELMo; Peters et al., 2018) can explain more of the variance in the data, but only modestly, only for stimuli that contain informative context words, and only when the scalar terms themselves are sufficiently context-sensitive.