Franz Schmalhofer

A Unified Model for Predictive and Bridging Inferences

Text materials such as those introduced by McKoon and Ratcliff (1986) have been repeatedly used to shape a theoretical understanding of inference processes. Recent results of Keefe and McDaniel (1993) with these materials were intriguing with respect to the generation and persistence of predictive and bridging inferences. To account for these data, the authors developed a formal model within Kintschs (1988, 1998) construction-integration theory. Computer simulations confirmed that the model explains the data well. The models key features are that (a) inferences may be generated and represented at the situational level and thereby differ from explicit statements, which may be encoded in a bottom-up fashion at the surface, propositional, and situational levels and (b) the maintenance of inferences and explicit statements depends on the interconnectivity of the multilevel representation rather than on an independent strength value of an individual knowledge unit. Recent data are described that support these theoretical assumptions. On the basis of this theoretical and empirical work, a unified model is proposed for the generation and persistence of predictive and bridging inferences. The implications of this unified model relative to previous theories are discussed, and a general taxonomy of inference processes is outlined.

Event-Related Potential Indicators of Text Integration across Sentence Boundaries.

An event-related potentials (ERPs) study examined word-to-text integration processes across sentence boundaries. In a two-sentence passage, the accessibility of a referent for the first content word of the second sentence (the target word) was varied by the wording of the first sentence in one of the following ways: lexically (explicitly using a form of the target word); conceptually (using a paraphrase of the target word), and situationally (encouraging an inference concerning the referent of the target word). A baseline condition had no coreference between the two sentences. ERP results on the target word indicated multiple effects related to word identification and word-to-referent mapping processes. Both the explicit and paraphrase conditions, but not the inference condition, showed a reduced N400 relative to the baseline condition, consistent with immediate integration by lexico-semantic processes. A 300-ms effect (P300) was found in the paraphrase condition. The results were consistent with an immediate integration hypothesis and furthermore differentiated a lexical (N200), a conceptual (P300), and a situational (N400) component for this integration. The conceptual basis appears not to extend to predictive inferences.

Neural indicators of inference processes in text comprehension: An event-related functional magnetic resonance imaging study

During language comprehension, readers or listeners routinely infer information that has not been stated literally in a given text or utterance in order to construct a coherent mental representation (situation model). We used a verification task in a behavioral study and in an event-related functional magnetic resonance imaging (fMRI) experiment to investigate the inference construction process. After having read sentences that mention the outcome of an event explicitly, implicitly, or not at all, participants verified the plausibility of short statements with respect to the context of the just read sentence. The results of the behavioral study established the verification task as a valid method for studying inferences. In the fMRI study, the dorsomedial prefrontal cortex was the most prominent area that was involved in the processing of inference statements. Regions in the left and right temporal lobes were associated with comparison processes that are based on the propositional representations of context sentence and test statements. The results are discussed with respect to levels of representations and the memory systems that underlie the verification process in the different sentence conditions.

Knowledge Mediation: A Procedure for the Cooperative Construction of Domain Ontologies

Up to now, there are few detailed proposals for the cooperative (and distributed) construction of ontologies (cf. [2]). The problem of how to establish a consensus and a shared conceptualization, especially when dealing with contradictory knowledge and conflicting interests has hardly been dealt with. We propose and evaluate a three-phased knowledge mediation procedure which is especially conceived to integrate different perspectives and information needs into one consensual ontology.

What is minimal about predictive inferences

Pronunciation of a probe word that relates to a highly predictable event is typically facilitated when the probe immediately follows the target sentence, but not when it is delayed until after a second sentence. These findings suggest that inferences about highly predictable events are minimally drawn during reading. Using a similar paradigm, in which the probe was delayed, we found significant priming in a reading condition that encouraged focus on the referential state of affairs (situation level) expressed in the sentences. In contrast, significant priming forpredictive texts was not observed in a reading condition that encouraged focus on the text per se (text level). These results are interpreted within a theoretical model that assumes that predictive inferences are represented at a situational level only, whereas explicitly presented information is represented at several levels (text and situational). It is in this sense that predictive inferences may be said to be minimally represented.

Strategic influence on the time course of predictive inferences in reading

In the present study, we investigated how reading strategies affect the time course of online predictive inferences. Participants read sentences under instructions either to anticipate the outcomes of described events or to understand the sentences. These were followed by a target word to be named, with stimulus onset asynchronies (SOAs) of 500 or 1,000 msec (50- or 550-msec interstimulus interval, respectively). Sentences either were predictive of events or were lexically matched control sentences. There was facilitation in naming latencies for predictable target words in the strategic-anticipation condition at both SOAs, but not in the read-to-understand condition, with a significant improvement in the former condition in comparison with the latter. This suggests that predictive inferences, which are typically considered to be resource demanding, can be speeded up by specific goals in reading. Moreover, this can occur at no cost to comprehension of explicit information, as was revealed by a comprehension test.