Neural correlates of quantity processing of numeral classifiers.

Objective: Classifiers play an important role in describing the quantity information of objects. Few studies have been conducted to investigate the brain organization for quantity processing of classifiers. In the current study, we investigated whether activation of numeral classifiers was specific to the bilateral inferior parietal areas, which are believed to process numerical magnitude. Method: Using functional MRI, we explored the neural correlates of numeral classifiers, as compared with those of numbers, dot arrays, and nonquantity words (i.e., tool nouns). Results: Our results showed that numeral classifiers and tool nouns elicited greater activation in the left inferior frontal lobule and left middle temporal gyrus than did numbers and dot arrays, but numbers and dot arrays had greater activation in the middle frontal gyrus, precuneus, and the superior and inferior parietal lobule in the right hemisphere. No differences were found between numeral classifiers and tool nouns. Conclusion: The results suggest that quantity processing of numeral classifiers is independent of that of numbers and dot arrays, supporting the notation-dependent hypothesis of quantity processing.

Although most of the studies on the processing of quantity information have typically focused on numerical processing (symbolic or nonsymbolic processing), some studies have investigated the neural correlates of quantity processing in natural languages. For example, quantifiers have attracted some attention from researchers (e.g., Cappelletti, Butterworth, & Kopelman, 2006;Cipolotti, Butterworth, & Denes, 1991;McMillan, Clark, Moore, Devita, & Grossman, 2005;McMillan, Clark, Moore, & Grossman, 2006;Morgan et al., 2011;Polk, Reed, Keenan, Hogarth, & Anderson, 2001;Troiani, Clark, & Grossman, 2011;Troiani, Peelle, Clark, & Grossman, 2009;Wei, Chen, Yang, Zhang, & Zhou, 2012). Quantifiers are carriers of quantity information in natural languages, and their meanings depend on "mapping a truth-value to a set of objects or to a quantity of a mass substance" (Morgan et al., 2011, p. 3532). There are various types of qualifiers (see Appendix A for details of quantifiers that have been studied in brain imaging and neuropsychological research).
Thus far, results of the studies on the neural basis of quantifier processing have been inconsistent. Some studies have shown that quantity processing of quantifiers shared a neural basis similar to that of numbers. Both of them appear to be supported by the bilateral IPS. Researchers have reasoned that the comprehension of quantifiers requires a numerosity device (e.g., McMillan et al., 2005McMillan et al., , 2006Morgan et al., 2011;Troiani et al., 2009Troiani et al., , 2011 and knowledge about numbers (McMillan et al., 2006) and therefore should activate the parietal cortex. Consistent with that argument, Polk et al. (2001) reported that their patient who had lesions in the left parietal lobe was deprived of abilities to process symbolic numbers and numerosity-related words (e.g., dozen, half, pair, single, and quarter), but possessed semantic knowledge about other kinds of words. Another report showed that a patient with semantic dementia preserved the ability of quantifier processing (e.g., several apples, a lot of cars) and nonverbal number processing, but showed impaired nonquantity processing (e.g., the girl with blond hair), suggesting that quantifiers and numbers share the same semantic system as numerical concepts (Cappelletti et al., 2006). A study of corticobasal syndrome patients and posterior cortical atrophy patients found that their impairment in processing cardinal quantifiers (e.g., "at least three"; see Appendix A for definitions of the types of quantifiers and additional examples) was partly due to a deficit in quantity knowledge related to temporal-parietal atrophy (Morgan et al., 2011). Finally, a neuroimaging study also showed that both first-order (e.g., "at least three") and higher order quantifiers (e.g., "less than half") activated the right inferior parietal cortex (McMillan et al., 2005).
Several studies further explored different types of quantifiers in detail, and found that the different types of quantifiers showed distinct neural bases involving a large-scale neural network including the parietal, frontal, and temporal cortices (e.g., McMillan et al., 2006;Morgan et al., 2011;Troiani et al., 2009Troiani et al., , 2011. McMillan et al. (2006) suggested that first-order quantifiers (e.g., "all," "some," and "at least three") involve number knowledge such as the numeric property of a set and associated brain regions such as the IPS, whereas higher order quantifiers (e.g., "less than half," "odd number of," and "even number of") involve comparative judgment (and hence working memory and associated brain regions) as well as number knowledge. Similarly, Morgan et al. (2011) argued that an impairment in the processing of cardinal quantifiers (e.g., "at least three") is partly due to a deficit in quantity knowledge related to temporal-parietal atrophy, whereas the impairment in the processing of logical quantifiers (e.g., "some") and majority quantifiers (e.g., "at least half") is associated with the ability of perceptual logic comprehension or executive functions subserved by the frontal cortex. Finally, Troiani et al. (2009) also showed that cardinal quantifiers (e.g., "at least three," "more than two," "even," and "odd") require magnitude processing, which depends on a lateral parietal-dorsolateral prefrontal network, and logical quantifiers (e.g., "all" and "some") require perceptual logic processing, which depend on a rostral medial prefrontal-posterior cingulate network.
However, other studies found a dissociation in neural bases between quantifiers and numbers. Wei et al. (2012) reported that nonnumerical quantifiers (e.g., logical quantifiers, such as "some" and "none") depend on the temporal and frontal cortex, but not the IPS, suggesting that the neural basis of quantifier processing (at least nonnumerical quantifiers) may be similar to that for general semantic processing, rather than that for numerical processing. Similarly, another type of quantifiers-the classifiers, a linguistic device used to express quantity information of measurement units of objects in natural languages (Her & Hsieh, 2010;Hwang, Yoon, & Kwon, 2008)-has been found not to rely on the IPS. In their study of a Gerstmann's syndrome patient with hypodensity in the left frontoparietal region, Cipolotti et al. (1991) found that the patient had severe impairment for number processing but was able to perform the task of deciding which of two  This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
classifiers (e.g., gram or kilo, meter or centimeter) expressed greater quantity. Therefore, it seems that classifier processing does not depend on number processing. The current experiment is the first functional MRI (fMRI) study focusing on numeral classifiers. It aimed to investigate the neural correlates of numeral classifiers among healthy subjects. As Appendix A shows, numeral classifiers are a type of quantifiers with explicit quantity information that "must occur with a number and/or a demonstrative, or certain quantifiers before a noun" (C. Li & Thompson, 1981, p. 104). In Chinese language, numeral classifiers are syntactically obligatory in counting the quantity of head nouns, in which the numerals denote the numerical concepts of a set of objects, especially the cardinality of the set, and the classifiers describe the quantity properties or imputed characteristic of the objects as measuring units (P. Li, Barner, & Huang, 2008;Zhang, 2007). Note. Clusters that survived p Ͻ .01 (uncorrected) with spatial extent k Ͼ 50 voxels were considered statistically significant. This document is copyrighted by the American Psychological Association or one of its allied publishers. This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
In this study, numeral classifiers were compared with numbers, dot arrays, and nonquantity words (i.e., tool nouns). Based on previous results on numeral classifiers and quantifiers (e.g., Cipolotti et al., 1991;Wei et al., 2012), we hypothesized that numeral classifiers would not activate functional regions specialized for quantity processing of numbers (i.e., the IPS), but rather would activate the same regions as those for tool nouns. The latter was based on the close relations between numeral quantifiers and nouns: The classifiers would limit the types of head nouns presented after them by determining properties (e.g., animacy, function, size, shape, and consistence) associated with objects denoted by the head nouns (e.g., P. Li et al., 2008;Zhang, 2007). If our hypothesis was confirmed, our results would suggest that quantity processing depends on type of symbols representing quantities-the notation-dependent hy-pothesis of quantity processing (e.g., Campbell & Clark, 1988;Eger et al., 2003).

Method Participants
Eighteen native Mandarin speakers were recruited from Beijing Normal University (BNU; nine women, mean age ϭ 22.4 years, SD ϭ 2.56). All participants were right-handed based on the Edinburgh Handedness Inventory (Oldfield, 1971). They had normal or corrected-to-normal visual acuity and no history of psychiatric or neurological disorders. They completed a screening form required by the BNU Imaging Center for Brain Research to ensure image quality and participants' safety. In- Figure 4. The contrasts of numeral classifiers with tool nouns, numbers, and dot arrays. Height threshold: p Ͻ .001, uncorrected. Extent threshold: k ϭ 10 voxels. Voxel size: 3 ϫ 3 ϫ 3 mm 3 . There were no differences between classifiers and tool nouns (upper panel). There were activation differences between classifiers and numbers (middle panel) and between classifiers and dot arrays (bottom panel). Greater activation for classifiers was typically in the left inferior frontal gyrus and left middle temporal gyrus (light gray), and greater activation for numbers or dot arrays was typically in the right inferior parietal lobule (dark gray). This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
formed consents were obtained following the protocol approved by the Institutional Review Board of the BNU Imaging Center.

Stimuli and Materials
There were four experimental conditions: numeral classifiers, tool nouns, numbers, and dot arrays. Fifty-eight numeral classifiers and 58 tool nouns were used (see Appendix B). These numeral classifier phrases included only a numeral (the number "one") and a classifier to reduce influence of other types of words. The numeral (the number "one") ensured that the participants would treat these phrases as classifiers rather than their alternative meanings. For the number condition, 59 two-digit numbers were chosen within the range from 10 to 99; for the dot condition, 59 dot arrays had quantities that also ranged from 10 to 99. The number or word stimuli were 30 mm in height and 15 mm in width, and the size of the box for the dot arrays was 60 mm in height and 60 mm in width. Note. Region clusters that survived p Ͻ .001 (uncorrected) with spatial extent k Ͼ 10 voxels were considered statistically significant. This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

Procedure
We used a block design for the semantic distance comparison task because such a design has been found to have superior statistical power (Friston, Zarahn, Josephs, Henson, & Dale, 1999). There were two 8-min runs. In total, 144 trials were divided equally into 16 blocks, eight for each of the two runs. Each block lasted 36 s, and was followed by a 24-s rest, with no repeated targets in the same block (see Figure 1). During the experiment, the order of the blocks was pseudorandomized, and the trial order was randomized within each block for each paorticipant. The order of the blocks was counterbalanced by dividing the 18 participants into two groups, each with a different order of the blocks. Before scanning, participants completed a practice block consisting of nine trials.
In each trial of the task blocks, the stimulus was presented in the center of the screen for up to 3,500 ms, and participants were asked to give a response as quickly as possible. Once participants responded, the stimulus disappeared, followed by a blank interval to complement the remaining time; 500 ms later, the next trial appeared. Each task block was followed by a rest block of 24 s, during which participants were asked to view a fixation cross. The stimuli were presented via back-projection onto a semilucent screen and then reflected to a mirror attached to the head coil. Exposure and timing of stimuli were controlled by Eprime software.
A semantic distance comparison task was used for all four types of stimuli (numeral classifiers, numbers, dot arrays, and tool nouns). A target item was presented in the top part of the screen and two alternative items at the bottom horizontally. Participants were instructed to judge which of the two items at the bottom had a closer semantic relation with the target. Participants pressed a key with the hand corresponding to the position of the answer.

Statistical Analysis of the fMRI Data
The preprocessing stage, including realignment, normalization, and smoothing, was performed using SPM 5 (http://www.fil.ion .ucl.ac.uk/spm/software/spm5/). All volumes were realigned to the first volume and spatially normalized to a common value to correct for whole brain differences over time. Data were normalized to a standard template in the Montreal Neurological Institute space for spin history and then smoothed with an isotropic 4-mm full-widthhalf-maximum Gaussian kernel and high-pass filter at a cutoff of 128 s. The images of the 18 participants were entered into a two-step statistical analysis to examine the activation patterns elicited by numeral classifiers in contrast to the other three types of stimuli. Figure 5. The contrasts of tool nouns and other conditions. Height threshold: p Ͻ .001, uncorrected. Extent threshold: k ϭ 10 voxels. Voxel size: 3 ϫ 3 ϫ 3 mm 3 . There were activation differences between tool nouns and numbers (upper panel) and between tool nouns and dot arrays (bottom panel). Greater activation for tool nouns was typically in the left inferior frontal gyrus (light gray), and greater activation for numbers or dot arrays was typically in the right superior and inferior parietal lobules (dark gray). This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly. Figure 2 shows the mean reaction time (RT) and error rates. RTs and error rates were analyzed with a repeated measures analysis of variance (four types of materials: numeral classifiers, tool nouns, numbers, and dot arrays). The main effect of type of materials was significant in error rates, F(1, 18) ϭ 3.64, p Ͻ .05, but not in RT. Further simple effects tests showed that error rates were higher for dot arrays than for numeral classifiers, p Ͻ .01, and tool nouns and numbers, p Ͻ .001. Figure 3 and Table 1 present the brain activities for numeral classifiers and the other types of materials (tool nouns, numbers, and dot arrays) relative to rest according to a conjunction analysis. Generally, the four conditions commonly activated the right inferior frontal gyrus, right angular gyrus, right supplementary motor area, right precentral gyrus, left insula, left cerebellum, and bilateral lenticular nucleus.

Results
The contrast analysis between numeral classifiers and tool nouns showed no areas with significant activations, suggesting that these two types of materials had similar neural basis. The effect size for each condition in each region is shown in Figure 4, upper panel, and Table 2 (p Ͻ .001, uncorrected, and spatial extent k Ͼ 10 voxels). Even with a more lenient threshold (p Ͻ .01, uncorrected, and spatial extent k Ͼ 10 voxels), the results did not change. Note. Region clusters that survived p Ͻ .001 (uncorrected) with spatial extent k Ͼ 10 voxels were considered statistically significant. This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
The comparison between numeral classifiers and numbers revealed that numeral classifiers had greater activation in the inferior frontal gyrus and middle temporal gyrus in the left hemisphere, but numbers induced more activation in the precuneus in the left hemisphere, and the superior and middle frontal gyrus, superior and inferior parietal lobule, precuneus, angular gyrus, and cerebellum in the right hemisphere. The effect size for each condition in each region is plotted in Figure 4, middle panel, and Table 2.
The comparison between numeral classifiers and dot arrays found that numeral classifiers elicited higher activation in the inferior frontal gyrus in the left hemisphere. Dot arrays elicited more activation in the middle frontal gyrus, cingulate gyri, superior occipital gyrus, putamen, and calcarine fissure in the left hemisphere, and the cingulate gyri, lingual gyrus, fusiform gyrus, inferior parietal lobule, superior frontal gyrus, and precentral gyrus in the right hemisphere (see Figure 4, bottom panel, and Table 2).
The contrast analysis between tool nouns and numbers showed that tool nouns had stronger effects only in the left inferior frontal gyrus; numbers elicited stronger activation in the superior and middle frontal gyrus, inferior parietal lobule, and precuneus in the left hemisphere, and the superior and inferior parietal lobule, superior and middle frontal gyrus, orbital part of inferior frontal gyrus, inferior temporal gyrus, middle occipital gyrus, and precuneus in the right hemisphere (see Figure 5, upper panel, and Table 3).
The contrast analysis between tool nouns and dot arrays showed that tool nouns elicited more activation in the left inferior frontal gyrus; dot arrays induced more activation in the cingulate gyri in the left hemisphere, and the superior and inferior parietal lobule, middle frontal gyrus, orbital part of inferior frontal gyrus, superior and middle occipital gyrus, fusiform gyrus, cingulate gyri, precuneus, and insula in the right hemisphere (see Figure 5, bottom panel, and Table 3).
The contrast analysis between numbers and dot arrays found that numbers did not elicit significantly more activation than dot arrays, and dot arrays had stronger activation only in the right fusiform gyrus (see Figure 6).

Discussion
We used fMRI to investigate the neural basis of processing numeral classifiers, and found that all four types of stimuli (numeral classifiers, numbers, tool nouns, and dot arrays) activated the inferior frontal gyrus, angular gyrus, supplementary motor area, and precentral gyrus in the right hemisphere, the insula and cerebellum in the left hemisphere, and bilateral lenticular nucleus. Contrast analyses showed that classifiers had similar activation with tool nouns, which was distinct from activations elicited by numbers and dot arrays. Classifiers induced more activation in the left frontal and temporal areas, and numbers showed stronger activation in the left precuneus and the right frontal and parietal areas. These findings suggest that unlike numbers or dot arrays, numeral classifiers do not rely on the IPS. In other words, the IPS does not seem to be a brain region for all abstract quantity processing. This notion is consistent with the growing literature that semantic processing is supported by distributed networks (e.g., see reviews by Cappa, 2012;Price, 2012).
First, our results that numbers and numerosities activated the parietal cortex are consistent with previous studies (see a metaanalysis by Arsalidou & Taylor, 2011;reviews by Cantlon et al., 2009;Cohen Kadosh et al., 2008;Nieder & Dehaene, 2009). The bilateral inferior parietal area is considered an important region for quantity processing and is independent of input-output channels and tasks (e.g., Chochon, Cohen, van de Moortele, & Dehaene, 1999). In addition, we also found that numbers and dot arrays had greater activation in the prefrontal cortex than did numeral classifiers and tool nouns. This was probably due to the involvement of visuospatial working memory in number and numerosity processing (e.g., Holloway, Price, & Ansari, 2010, for Arabic numbers and squares; Jacob & Nieder, 2009, for fractions).
Second, the brain regions that were activated more strongly by classifiers and tool nouns than by numbers and dot arrays were the classical areas for general semantic and verbal processing, especially in the left hemisphere (Booth et al., 2006). For example, in an early PET study of word processing, Martin, Wiggs, Ungerleider, and Haxby (1996) found that passive viewing of tool nouns induced activation in the left inferior frontal gyrus and middle temporal gyrus. Later research has confirmed the role of the left inferior frontal gyrus and left fusiform gyrus in the processing of words in different languages, including Korean (e.g., Yoon, Chung, Kim, Song, & Park, 2006) and Chinese (e.g., Xue, Dong, Jin, & Chen, 2004). Therefore, the quantity information in classifiers seems to be processed in the same brain regions as general semantic processing. Figure 6. The contrasts of numbers and dot arrays. Height threshold: p Ͻ .001, uncorrected. Extent threshold: k ϭ 10 voxels. Voxel size: 3 ϫ 3 ϫ 3 mm 3 . There were no greater activation for numbers than dot arrays, and the only area with greater activation for dot arrays than numbers was in the right fusiform gyrus (dark gray). This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
Our results add to a small but growing literature on the neural basis of quantifier processing. They are consistent with at least two previous studies. In a patient study, Cipolotti et al. (1991) found that the patient had impaired number processing but intact classifier processing. Wei et al. (2012) showed that quantifier processing did not rely on the IPS, but rather on the temporal and frontal cortices. Taken together, from the results of these studies as well as those of the current study, it appears that the processing of quantifiers, including classifiers, shares a common neural basis with language processing but not with number processing. These results also support the hypothesis that quantity processing is notation-dependent, not notationindependent (e.g., Dehaene, 1992). The notation-dependent hypothesis was proposed by Campbell based on results from a number of behavioral studies that showed how the presentation format (Arabic digits vs. English number words) influenced number processing such as estimation of numerical magnitude (Campbell, 1994;Campbell & Clark, 1988). It has been confirmed by a series of cognitive neuroscience experiments, which found that different types of numerical symbols induce different activation in the bilateral IPS (e.g., Ansari, 2007;Cohen Kadosh, Muggleton, Silvanto, & Walsh, 2010;Diester & Nieder, 2007;Holloway et al., 2010;Jacob & Nieder, 2009;Santens, Roggeman, Fias, & Verguts, 2010).