Modeling Concept Activation in Working Memory during Online Sentence Processing

There have been several computational alternatives to the cloze task (Taylor, 1953) intended to approximate word predictability effects on eye movements during reading. In this study, we implement a computational model that instantiates each content word in a sentence as an input that activates semantic concepts in working memory. The predictability of a word is then determined by the extent to which its corresponding semantic representation is associated with the network of concepts already active in working memory from the preceding context. The computation of concept activation is based on a connectionist model (Landscape model, see van den Broek, 2010). Latent semantic analysis (LSA) is used to establish connections between words and simulate the long-term semantic associations among concepts (Landauer & Dumais, 1997). This model provides a means of investigating how language comprehension is affected by the activation of concepts in working memory.