This talk presents a process model of human sentence comprehension based on the following theoretical claim: Sentence comprehension can be understood as a series of cue-based retrievals from short-term (and long-term) memory. Conceiving of sentence processing in this way lets us incorporate important ideas from general memory research in cognitive psychology, including similarity-based retrieval interference, activation decay, and confusable position codes for serial order information. These principles provide explanatory accounts of many parsing phenomena (such as difficulty on embeddings and recency effects), and generate novel predictions which can be empirically tested. I'll present some new data from reading experiments in English showing an interesting cross-over effect of interference and decay on attachment (the process of incorporating a new word into an existing partial interpretation) and reanalysis (recovering from a misinterpretation of a local ambiguity): attachment processes show strong effects of interference, but small effects of decay, while reanalysis shows just the opposite pattern. The process model is based on a computational cognitive architecture, ACT-R, which embodies a set of independently motivated hypotheses about memory retrieval and cognitive skill. The model yields the kind of detailed behavioral trace required to adequately bring theory into contact with temporally-rich paradigms such as eye-tracking.

This is joint work with Julie Van Dyke, JJ Nakayama, and Shravan Vasishth.