In this dissertation I describe a cognitive model of sentence processing. The model operates at the semantic level and can apply to verification or comprehension of metaphoric or literal sentences, isolated or embedded in discourse. It uses an incremental search--and--match mechanism to find a long-term--memory referent (interpretation) for an input sentence. The search is guided by cues such as the last few words read or previous tentative interpretations. The process of comprehension produces a propositional representation for the input sentence and also keeps track of local comprehension failures.

The model is implemented in the ACT-R framework and offers a scalable solution to the problem of language comprehension: its performance (in terms of speed and accuracy) is roughly invariant to the number of facts held in the long-term memory. Its predictions match data from psycholinguistic studies with human subjects. Specifically, the sentence-processing model can simulate the comprehension and verification of metaphoric and literal sentences, metaphor-position effects on sentence comprehension, semantic illusions and their dependence on semantic similarity between the distortion and the undistorted term. The products of the sentence-processing model can explain the pattern of sentence recall in text-memory experiments.

This dissertation also explores the modeling alternatives faced by the design of a sentence-processing model. I show that, to achieve comprehension speed comparable to that of humans, a model must minimize the explicit search process and rely on semantic associations among words. I also investigate how the representation chosen for propositions and meanings affects the comprehension process in a production-system framework such as ACT-R.

151 pages

Return to: SCS Technical Report Collection
School of Computer Science homepage

This page maintained by reports@cs.cmu.edu