Computational model reveals two sources of polarity across negative adjectival expressions and quantifiers

The polarity effect is one of the most robust effects in linguistics: Negative expressions are processed slower than corresponding positive ones. Decades of research has shown that this effect replicates for different types of negative expressions, however, its magnitude varies depending on the expression type. It has been argued that in addition to polarity, the downward monotonicity increases the cost of processing negative expressions. In this paper, we use computational modeling of data from a sentence-picture verification task to investigate the sources of the polarity effect by contrasting quantifiers ('more than half' vs. 'fewer than half') that differ in polarity and monotonicity with adjectives ('a large proportion' vs. 'a small proportion') that only differ in polarity. We collected reaction times and response data in two web-based sentence-picture verification task experiments. Our reaction time analysis showed a larger polarity effect for quantifiers than adjectives, thus replicating a previously observed non cross-over interaction between polarity and expression type. In order to interpret this interaction with respect to semantic properties of these expressions, we fit scalinDDM, a modified version of the Diffusion Decision Model to the data. We showed that the drift rate and non-decision time parameters were affected by polarity, however, the effect in drift rate was larger for quantifiers than adjectives. This finding suggests there is no exclusive mapping between semantic properties and the model's parameters. Our results show that the computational model can increase transparency in mapping between linguistic properties and dependent variables in an experiment.