Predictability versus activation of representations: Evidence from sentence comprehension
MetadataShow full item record
The amount of processing difficulty a reader has in integrating a new word into a sentence is a function of the degree to which information about that word has been activated by context, prior to that word being encountered. Some computational models of language comprehension, like the surprisal models (Hale, 2001; Levy 2008), have attempted to capture the relationship between how activated information associated with a target word is, prior to its being encountered, and how easy or hard that target word will be to integrate into a sentence representation, by assuming that the activation of an about-to-be-encountered word is completely determined by how predictable it is given a sentence context. In this dissertation, I tested the claims of the probability-only account by showing that word predictability did not entirely determine the degree of processing difficulty. I present three results that the probability-only account cannot explain. The first result is that the costs of integrating a target word into a context sentence were not exclusively predicted by its conditional probability. The degree to which target words were semantically similar to other possible words for the same context also influenced the ease or difficulty of integrating that word into a sentence (Chapter 3). Second, facilitation based on semantic similarity emerged only when contexts weakly constrained the properties of possible word choices (Chapter 4). Third, the effect of semantic similarity on word integration was modulated by whether or not verbs imposed early and highly selective constraints on the activation of features associated with instrument participant roles (Chapter 5). All of these results indicate that the activation of semantic concepts or features from context is not entirely reflected in word predictabilities. In contrast to probability-only accounts, I argue that how easy or hard it is to integrate a word into a sentence is best understood in terms of the activation of representations across word and semantic feature levels.