Involvement of left inferior frontal gyrus in sentence-level semantic integration

Abstract

Using event-related functional MRI, we examined the involvement of the left inferior frontal gyrus (LIFG) in semantic integration in reading Chinese sentences. During scanning, Chinese readers read individually presented sentences and judged whether or not a sentence was semantically acceptable. Behaviorally, those sentences with a small degree of semantic violation were found to be more difficult to reject relative to sentences with a large degree of semantic violation, indicating that more semantic integration occurred in the former than in the latter condition. Direct contrast revealed significantly greater brain activity in the LIFG for sentences with a small violation, relative to those with a large violation, but no differences in any anterior temporal cortical areas between the two types of anomalous sentences. The results are in line with the idea that the LIFG plays a critical role in integrating individual word meanings to coherent sentence-level messages, but not with the idea that semantic integration depends on anterior temporal cortex in language comprehension.

Introduction

In recent years, there has been an increasing interest in language processing and its neural correlates (e.g., Constable et al., 2004, Hagoort, 2003, Hagoort, 2005, Hickok and Poeppel, 2004, Jung-Beeman, 2005, Spitsyna et al., 2006, Wang et al., 2008, Willems et al., 2008). As a result, different models have been postulated on the basis of the findings in cognitive neuroscience research.

A major model on the topic is the Bilateral Activation, Integration, and Selection (BAIS) model, which proposes that semantic processing in language comprehension consists of three distinct but highly interactive components (i.e., activation, selection, and integration), each of which is supported by different brain areas (Jung-Beeman, 2005). Specifically, bilateral Wernicke's areas are considered to be responsible for activating or retrieving lexical and semantic information, the inferior frontal gyrus (IFG) for selection among competing semantic activations, and anterior temporal areas for semantic integration to compute semantic overlap among multiple semantic activations and to construct higher-order semantic relations.

Another major model on the same topic is the one proposed by Hagoort and colleagues (Hagoort, 2003, Hagoort, 2005, Hagoort and van Berkum, 2007), in which they argue that a different set of three major components (i.e., memory, unification, and control) is crucial for language comprehension and production. According to this Memory, Unification, and Control (MUC) model, the left temporal cortex underlies the memory component in retrieving word information from long-term memory, the LIFG underlies the unification component in combining retrieved information into larger units, and the anterior cingulate cortex and dorsolateral prefrontal cortex (DLPFC) underlie the control component in planning verbal actions and allocating attention.

Semantic integration is a key component in the construction of sentence-level meaning representations, but different researchers attribute this process to different brain areas (i.e., anterior temporal cortex vs. LIFG). To examine which one of the two mentioned cortical areas is connected with semantic integration in language processing, the present study adopted a violation paradigm commonly used for studying semantic processing (e.g., Dapretto and Bookheimer, 1999, Hagoort et al., 2004, Kutas and Hillyard, 1980; see review in Kaan and Swaab (2002) and Kutas et al. (2006)). The violation paradigm typically involves comparison of normal sentences with anomalous sentences that have been violated in a certain manner (e.g., a semantic violation in the present study). The basic assumption is that participants are subject to increasing demand on and recruit more cognitive resources for integration in reading anomalous sentences. Following the same logic, brain regions more strongly activated by the anomalous sentences relative to the normal sentences are generally interpreted as possible neural correlates of a particular type of reading related operations (e.g., Baumgaertner et al., 2002, Hagoort et al., 2004, Homae et al., 2002).

It is important to note, however, that there are two potential problems in using the above mentioned violation paradigm to investigate semantic integration in language processing. First, normal sentences and semantically violated sentences differ not only in semantic integration, but also in other aspects. For example, anomalous sentences may recruit processes not engaged by normal sentences at all, such as violation detection and violation repairing (Indefrey et al., 2001, Kaan and Swaab, 2003). These processes, though unrelated to semantic integration, may also show up in the anomalous vs. normal comparison. Although similar problems have been recognized in previous syntactic violation studies (e.g., Indefrey et al., 2001), no attempt has been made so far to deal with it in semantic studies.

Another potential problem is that, when comparing the anomalous and normal sentences, the two types of sentences are actually associated with different responses. In fact, in making semantically acceptable or unacceptable judgments, participants are essentially matching the constructed sentence meaning with some previously established, semantic representations so that semantically acceptable sentences require a match or positive response and semantically unacceptable sentences require a mismatch or negative response. It has been well established in cognitive psychology that positive and negative decisions may involve different cognitive processes (see, e.g., Farell, 1985, Sternberg, 1966) and Treisman and Gormican (1988) for evidence in behavioral studies; e.g., Zhang et al. (2003) for evidence in brain imagining studies), and that making positive and negative responses may be influenced by factors such as task demand, response bias, and response strategy (see review in Farell (1985)). Hence, comparison of anomalous and normal sentences may reflect general decision-making processes not specific to semantic integration. In other words, semantic integration is only one of several factors that may differentiate anomalous and normal conditions, if the two conditions require different responses. Under this situation, one should be cautious in interpreting brain activation differences across the two conditions.

In the present study, we attempted to make use of the principle of parametric design to compare two sentence types with different degrees of semantic violation. Table 1 illustrates the experimental conditions, each with a sample sentence. The anomalous sentences were constructed by changing a target noun in a normal sentence. For example, when the normal sentence is as follows: “The construction workers use pump to draw underground water” (English translation), the target word “pump” is replaced with “iron” to produce a relatively small semantic violation, and with “salt” to produce a relatively large semantic violation.

Although neither “iron” nor “salt” can be integrated with the rest of the sentence to produce a coherent meaning representation, the former word is semantically related to the original word “pump”, and is more likely to fit the sentence context, relative to the latter, which is semantically unrelated to the original word (see similar manipulations in Federmeier and Kutas (1999)). Consequently, we expected that it would be more difficult for participants to detect the violation in a small semantically violated sentence and reject the sentence as semantically acceptable, relative to a large semantically violated sentence. This is because participants would need to engage more semantic integration in order to make sense of the sentence in the small violation condition than in the large violation condition.

Relative to the anomalous vs. normal contrast, the small vs. large violation contrast would be much more comparable, because they involve the same processes of violation detection, semantic repair, and require the same negative response. Therefore, the contrast between the small vs. large violation should be more appropriate to reveal the neural correlates of semantic integration and to test the specific predictions from models such as the MUC and the BAIS.