Word order and lexical semantic context effects in sentence comprehension

AUTOCODER (Durbin, Earwood, & Golden, 2000) is a computer program employing heuristic shallow processing strategies to automatically identify propositions in free response data in order to support data analysis in experimental psychology. This thesis investigates whether the shallow sources of information used by AUTOCODER are also used by humans in sentence comprehension. Each of twenty participants read a context word sequence such as the father the son helped (noun-noun-verb) followed by a target word sequence such as the man the woman visited (noun-noun-verb). Participants were then asked to identify the agent of the target word sequence. Increasing the presence of either word-order cues and lexical semantic cues in the context word sequence was shown to differentially influence the degree to which a particular noun in the target word sequence would be assigned the agent role. Next, the Statistical Competition Model (SCM) was introduced which learns associations between role-assignments and word sequences. By manipulating its memory representation, the SCM model can be biased to increase the likelihood a noun is assigned the agent role. The results of these manipulations accounted for the key experimental findings of the effects of word order and lexical semantics on agent role assignment in humans. These manipulations also yielded three predictions. First, the magnitude of such bias effects is fundamentally graded, as opposed to "all or none". Second, the verb-noun-noun word sequence bias effect is greater than the noun-noun-verb word sequence bias effect, which in turn is greater than the noun-verb-noun bias effect. Third, overall preference for assigning the first noun to the noun-noun-verb condition is greater than the overall preference for assigning the first noun to the verb-noun-noun condition. In order to evaluate the computational adequacy of the SCM model, the extended version of SCM, AUTOCODER, was trained to identify propositions and their corresponding clause boundaries in summarization free response data from 70 second and fifth graders. AUTOCODER's performance in coding and identifying clause boundaries in new summarization data from 24 other students was comparable (92% agreement) to that of an experienced human coder, supporting the computational adequacy of the SCM model.