Naïve discrimination learning approach to polysemy

Abstract

In this paper, the focus of the investigation is the power of the computational model which is based on the principles of discrimination learning to predict processing latencies of polysemous nouns. Discrimination learning has been shown as a powerful principle that can account for multiple language phenomena, such as various frequency effects, various morphological phenomena and so on. However, the application of the model in predicting semantic effects is at its very roots. Here, the model is applied to simulate processing of words with multiple related senses, by mapping the bigrams from the input to the output that is represented by an array of words that co- occur with the target word. The simulated reaction time was positively correlated with empirically observed processing latencies. Additionally, the regression model fitted to simulated reaction time revealed a significant effect of lemma frequency, word familiarity, number of senses, and redundancy of sense probability distribution. These effects mirrored the pattern of the effects observed with empirical reaction time, the only exception being the reversal of the direction of the number of senses effect. Taken together, the results add to the body of research that links polysemy to context variability and additionally demonstrate how such a link can be related to learning.