How the brain processes causal inferences in text: A theoretical account of generation and integration component processes utilizing both cerebral hemispheres

Theoretical models of text processing, such as the construction-integration framework, pose fundamental questions about causal inference making that are not easily addressed by behavioral studies. In particular, a common result is that causal relatedness has a different effect on text reading times than on memory for the text: Whereas reading times increase linearly as causal relatedness decreases, memory for the text is best for events that are related by a moderate degree of causal relatedness and is poorer for events with low and high relatedness. Our functional magnetic resonance imaging study of the processing of two-sentence passages that varied in their degree of causal relatedness suggests that the inference process can be analyzed into two components, generation and integration, that are subserved by two large-scale cortical networks (a reasoning system in dorsolateral prefrontal cortex and the right-hemisphere language areas). These two cortical networks, which are distinguishable from the classical left-hemisphere language areas, approximately correspond to the two functional relations observed in the behavioral results. In order to fully understand a narrative text, readers must be able to mentally link together successive events to form a coherent representation of the story. Often, the events in the story are not explicitly related to each other, and the reader must connect them by generating and integrating inferences. Through the history of discourse-processing research, cognitive psychologists have struggled with the questions of under what circumstances and how such connecting causal inferences are generated. With the advent of functional magnetic resonance imaging (fMRI) techniques, it is now possible to integrate cognitive behavioral findings with brain-imaging research to inform answers to these questions by investigating the neural bases of the component processes involved in the generation and integration of inferences. The inference process requires that a reader first generate a possible inference and then attempt to integrate it into the internal representation of the text. The construction-integration (CI) model of text comprehension (Kintsch, 1988) is consistent with this general description of inferencing. According to the CI model, an initial process in which the reader will liberally generate many possible inferences is followed by a second process of integrating those inferences that have a high degree of connection with the text base into the representation of the text. A successful integration of an inference will then result in a text representation that involves both the specific propositions contained in the text and those inferred propositions that were generated by the reader to connect information in the text. Although various types of inferences have been studied and categorized (e.g., Singer, 1994; van den Broek, 1994), our study focused on causal inferences. Keenan, Baillet, and Brown (1984) and Myers, Shinjo, and Duffy (1987) investigated the relationship between memory for text and the degree of causal relatedness between the sentences within the text. They created sentence pairs that varied across four levels of intersentence causal relatedness. An ‘‘outcome’’ sentence, such as The next day his body was covered with bruises, was preceded by one of four different types of sentences describing antecedent conditions: highly related: Joey’s big brother punched him again and again. moderately related 1: Racing down the hill, Joey fell off his bike. moderately related 2: Joey’s crazy mother became furiously angry with him. distantly related: Joey went to a neighbor’s house to play. Myers et al. (1987) and Keenan et al. (1984) found that reading times increased as the relatedness of the sentences decreased. Somewhat counterintuitively, memory for the two-sentence passages (measured by various recall and recognition tests) followed an inverted-U-shaped function: The moderately related sentences were recalled (and recognized) better than the highly related or distantly related pairs. This is counterintuitive because the pattern is not a monotonic function of either reading time or causal relatedness. Address correspondence to Robert Mason, Center for Cognitive Brain Imaging, Department of Psychology, Carnegie Mellon University, Pittsburgh, PA 15213; e-mail: [email protected]. PSYCHOLOGICAL SCIENCE Volume 15—Number 1 1 Copyright r 2004 American Psychological Society Myers et al. (1987) and Myers and Duffy (1990) suggested that the increased recall for the intermediate causally related sentences was a result of the readers having generated a causal link between the two sentences. The generation of a causal link is presumed to produce a larger and perhaps richer interconnected network of nodes in the reader’s text-base representation, as shown in Figure 1. This larger, richer network can provide additional retrieval cues for recall, thus resulting in higher recall. We hypothesized that the reading of the moderately related sentence pairs was accompanied by both the generation and the integration of causal inferences. In the highly related sentences, this inferencing process was unnecessary, resulting in faster reading times and more sparse text representations. In contrast, the distantly related sentences had slower reading times, presumably as a result of a liberal generation of possible inferences to connect them, but lower recall, as a result of lack of success in integrating any inference. Using functional brain imaging, we attempted to find specific areas in the brain that respond differently to the two component processes of inference making in discourse comprehension. Studies have shown that fMRI is an excellent measure of the intensity of cognitive processing, specifically during reading (Just, Carpenter, Keller, Eddy, & Thulborn, 1996); this is true even when additional processing is not accompanied by an increase in reading times (Mason, Just, Keller, & Carpenter, in press). Consider the component process of generating possible inferences. If the linear increase in reading times with increasing causal distance is due to the generation of possible inferences, then there may exist a set of brain areas that shows a similar linear increase in brain activation. Similarly, if the inverted-U-shaped function relating recall to degree of causal relatedness is due to integration processes, there may exist a set of cortical areas involved in inference integration that shows a similar inverted-U function. Finally, it is quite likely that a set of brain areas involved with the basic levels of sentence processing (e.g., lexical access, syntactic parsing) will show equivalent activation across the levels of causal relatedness. It has been proposed that the right-hemisphere homologues of the left-hemisphere language areas (superior, middle, and inferior temporal gyri; inferior frontal gyrus, including pars opercularis and pars triangularis; and the inferior parietal area) are extensively utilized in discourse processing, particularly inference generation (for an extensive review, see Beeman, 1998). Evidence for the role of the right hemisphere in discourse processing has come from two lines of research. The first is neuropsychological investigations of patients with lesions who have trouble with aspects of discourse processing. Patients with lesions to the right hemisphere generally have trouble drawing inferences in order to integrate sentences and maintain coherence (Beeman, 1993; Brownell, Potter, Bihrle, & Gardner, 1986); they do not mistakenly recall inferences, presumably because they never generated them (Grafman, Salazar, Vance, Weingartner, & Amin, 1987); and they make elaborative inferences more easily than bridging inferences (Tompkins & Mateer, 1985). The second research area that has provided evidence for the right hemisphere’s involvement in discourse processing is experimental studies in which critical words from a text are presented separately to one of the two visual hemifields under the assumption that they will be processed first by the contralateral hemisphere. Beeman et al. (1994) have shown that when probes are inference related, they are primed in the left visual field (right hemisphere) immediately, and subsequently are primed in both the left visual field (right hemisphere) and the right visual field (left hemisphere). Additionally, there is a small set of brain-imaging studies that is beginning to illuminate brain function in discourse processing. Several positron emission tomography (PET) studies have examined discourse processing at a broad level; by comparing the comprehension of stories with the comprehension of unrelated sentences, Fletcher et al. (1995) and Mazoyer et al. (1993) were able to attribute activation in the left frontal gyrus to story processing. Bilateral inferior frontal and bilateral middle temporal gyri were activated when a moral judgment was required after reading a set of Aesop’s fables (Nichelli et al., 1995). St. George, Kutas, Martinez, and Sereno (1999), using fMRI, found greater right-hemisphere activation when the stories were not preceded by a title than when they were, but Maguire, Frith, and Morris (1999) did not find an increase in right-hemisphere activation in a similar task. Bottini et al. (1994) found several areas of righthemisphere activation during the processing of figurative language. Fig. 1. Possible representation of the reader’s internal network corresponding to each of the three types of causally related sentence pairs (reprinted from Myers & Duffy, 1990, with permission). 2 Volume 15—Number 1 Inferencing in the Brain