Semantic priming is a crucial phenomenon to study the organization of semantic memory. A novel type of priming effect, integrative priming, has been identified behaviorally, whereby a prime word facilitates recognition of a target word when the 2 concepts can be combined to form a unitary representation. We used both functional and anatomical imaging approaches to investigate the neural substrates supporting such integrative priming, and compare them with those in semantic priming. Similar behavioral priming effects for both semantic (Bread–Cake) and integrative conditions (Cherry–Cake) were observed when compared with an unrelated condition. However, a clearly dissociated brain response was observed between these 2 types of priming. The semantic-priming effect was localized to the posterior superior temporal and middle temporal gyrus. In contrast, the integrative-priming effect localized to the left anterior inferior frontal gyrus and left anterior temporal cortices. Furthermore, fiber tractography showed that the integrative-priming regions were connected via uncinate fasciculus fiber bundle forming an integrative circuit, whereas the semantic-priming regions connected to the posterior frontal cortex via separated pathways. The results point to dissociable neural pathways underlying the 2 distinct types of priming, illuminating the neural circuitry organization of semantic representation and integration.
Thousands of concepts are acquired in one's lifetime. Concepts are bound together through multiple relationships to form an interconnected conceptual network. The activation of one concept within the network can facilitate access to another concept if they are semantically similar (via feature overlapping, e.g., cookie–bread) or associated (generated by a free-association task, e.g., salt–pepper) (Collins and Loftus 1975; Hutchison 2003). A robust behavioral phenomenon, “lexical-semantic priming,” has been observed repeatedly, in which responding to a target word is faster when preceded by a semantically related prime word versus an unrelated one (Meyer and Schvaneveldt 1971). In addition, concepts can prime one another even if they are not already associated, but can be easily combined to form a unitary representation (e.g., Cherry – Cake), a phenomenon termed “integrative priming” (Estes and Jones 2009; Mather et al. 2014). The aim of the present study was to elaborate the potentially distinct neural underpinnings of semantic and integrative priming.
Integrative priming has been argued to reflect distinct cognitive processes from that of semantic priming (Estes and Jones 2009; Jones and Golonka 2012; Mather et al. 2014). In contrast to semantic priming, complimentary role assignment and conceptual combinatorial processes are implicated in integrative priming. According to the relational integration hypothesis (Estes and Jones 2009), if there is an integrative relationship in a word pair, a process of role assignment is automatically activated (Estes and Jones 2009; Mather et al. 2014) where the concepts are assigned to complementary semantic roles. For example, when participants see the pair Cherry – Cake with the words presented sequentially, the most common roles in which the Cherry is used to modify other concepts are activated. If Cake has a compatible dimension to be modified by Cherry (e.g., a cake has a particular type of flavor), this complementary semantic role can facilitate the lexical processing of the target word Cake, which is also the process involved when integrating individual concepts into a unique identity (Estes and Jones 2009). Behaviorally, it has been demonstrated that the integrative-priming effect tends to emerge in a short-time interval between a prime and a target, and diminish faster compared with that of in associative priming (Estes and Jones 2009). Moreover, integrative priming is not sensitive to the manipulation of relatedness proportion (the proportion of semantically related trials in a stimuli context) while semantic priming is (Estes and Jones 2009; Jones and Golonka 2012; Mather et al. 2014). These findings suggest that integrative priming and semantic priming involve different cognitive operations.
In parallel to the behavioral differences between the 2 types of priming effects, recent neuroimaging studies have demonstrated that integrative processes involved in both word and sentence comprehension trigger distinct patterns of neural activation from those of classical semantic priming. Specifically, a distributed frontal-temporal semantic network has been implicated in semantic priming (Hickok and Poeppel 2004; Lau et al. 2008; for meta-analysis, see Binder et al. 2009) both in terms of neural response suppression (Devlin et al. 2000; Gold et al. 2006; Lau et al. 2013) and enhancement (Kotz et al. 2002; Raposo et al. 2006; Sachs et al. 2011; Lee et al. 2014) relative to unrelated word pairs. In contrast, when experimental tasks involve conceptual combination, such as tasks focusing on basic integrative processing between adjectives and nouns (e.g., “red – boat”) (Graves et al. 2010; Bemis and Pylkkanen 2011, 2013) or on sentence-level comprehension (Lau et al. 2008; Rogalsky and Hickok 2009; Wilson et al. 2013), different regions were implicated including left anterior inferior frontal cortices, left anterior temporal cortices, bilateral tempo-parietal, and medial prefrontal regions. Nevertheless, the exact neuroanatomical bases of integrative priming, and whether integrative and semantic-priming effects are underpinned by different brain mechanisms, are still unknown.
To this end, the goal of the present study was to delineate the distinct neural substrates between integrative and semantic-priming operations. Specifically, we manipulated the relationship between prime and target words in a lexical decision task (LDT). Integrative word pairs were constructed that displayed high integrative potential but had a low level of prior association and were semantically dissimilar. In contrast, the semantically related word pairs were both highly similar and associated but had low integrative potential. These 2 types of word pairs were contrasted with a matched unrelated word pairs to define the priming effects. Such design permitted us to determine the neural substrate for integrative priming, while minimizing confounding factors from semantic similarity and association, and vice versa. In addition, to help isolate the neural underpinnings of the semantic or integrative priming, separate from processing involved in word recognition and access to the meanings of individual words (Badre et al. 2005), we constructed another condition in which the prime word was a meaningless nonword and the target word was the same as in the other conditions. Finally, we expect an increased activity pattern to integrative versus unrelated word pairs for the semantic integration regions mentioned above, reflecting distinct neural mechanism associated with the process of forming a new representation (Henson 2003; Lee et al. 2014).
Beyond the activation-based functional localization, we also investigated the brain circuitry underlying semantic and integrative priming. Researchers in both psycholinguistic and neurology of language have shown increasing interest in devising explicit models of the brain circuitry underpinning language functions. Indeed, the language network has been proposed as a highly interactive system (Dick and Tremblay 2012). Both activation of language-related regions and effective communication among them by fiber bundles are proposed to be associated with the implementation of language functions (Fedorenko and Thompson-Schill 2014). Using diffusion tensor imaging (DTI) technology, researchers can track the language-related fiber pathways in vivo (Dick and Tremblay 2012; Friederici 2012; Thiebaut de Schotten et al. 2012; Friederici and Gierhan 2013). Previous studies have found that several fiber bundles, such as the extreme capsule (EC) fiber system (ECFS), superior longitudinal fasciculus/arcuate fasciculus (SLF/AF), and uncinate fasciculus (UF), connecting the tempo-parietal cortices to the frontal cortices (Saur et al. 2008; Wong et al. 2011), largely contribute to human language processing (Friederici and Gierhan 2013). In the present study, we investigated the anatomical pathways associated with each type of priming by using a probabilistic fiber-tracking method to track the most likely fiber bundles directly connecting those priming-related regions. We hypothesized that the integrative-priming regions may connect with each other via specific fiber bundle(s) forming a circuit, while the semantic-priming regions may connect with each other by separated fiber bundle(s) forming another circuit.
Materials and Methods
A total of 28 native Chinese speakers participated in the experiment and were paid for participation. All participants were right-handed, with normal or corrected-to-normal vision, and no prior history of neuropsychiatric disorders. All participants signed a written consent form approved by the local ethical review board in South China Normal University. One participant was excluded due to poor task performance (accuracy <70%) in the LDT and another was excluded due to excessive head movements (>1 mm). Thus, the analyses were conducted on data from 26 participants (11 males; 18–28 years old, mean age = 21.2 years, standard deviation [SD] = 2.2).
Experimental Design and Stimulus Construction
To isolate neural substrates associated with integrative and semantic priming, our fMRI study adopted a priming paradigm in combination with a LDT. Two-character nouns in Chinese were used as primes and targets. Four types of prime–target relationships were constructed, including integrative (e.g., 樱桃/Cherry – 蛋糕/Cake), semantic (e.g., 面包/Bread – 蛋糕/Cake), unrelated (e.g., 司机/Driver – 蛋糕/Cake), and nonword (e.g., /Kmbol – 蛋糕/Cake). Here, the integrative (or semantic) priming effect was defined as the contrast of the integrative (or semantic) and unrelated condition. While the unrelated condition served as a control for assessing the impact of semantic or integrative relations, the nonword condition served as a control for dissociating lexical-level processes/effects (e.g., lexical-semantic retrieval of the meanings of the constituent words) from those beyond the single word level (e.g., priming processes). Because there was only one real word (i.e., the target, Cake in the example) in the nonword condition, the amount of meaning retrieval in the nonword condition was assumed to be less than other conditions. Thus, we defined the effect of word retrieval as the contrast of the unrelated and nonword conditions.
The materials included a total of 168 sets of words. Each set included a unique target word paired with 4 types of prime words. The integrative word pairs were constructed to have a high degree of integration potential (easy to combine prime and target into a meaningful phrase). Constituents of integrative pairs were selected to be semantically dissimilar and to have no prior association. In contrast, the semantically related word pairs were constructed to be both highly similar and associated but with minimal degree of integrative potential (Estes and Jones 2009; Mather et al. 2014). In the unrelated condition, the constituents were nonsensical if combined, semantically dissimilar and unassociated. In the nonword condition, a real-word target was paired with a nonword prime that did not convey any semantic information. These nonword primes were unpronounceable nonwords created by randomly assembling Chinese radicals (see Table 1 for more details).
|Integrative||樱桃 – 蛋糕|
[Cherry – Cake]
|4.77 (0.34)*||2.15 (0.58)||0.06 (0.14)*||0.02 (0.09)|
|Semantic||面包 – 蛋糕|
[Bread – Cake]
|2.19 (0.57)||4.04 (0.52)*||0.03 (0.07)||0.08 (0.15)*|
|Unrelated||司机 – 蛋糕|
[Driver – Cake]
|1.47 (0.49)||1.23 (0.21)||0 (0)||0 (0)|
|Nonword|| – 蛋糕|
[Kmbol – Cake]
|Integrative||樱桃 – 蛋糕|
[Cherry – Cake]
|4.77 (0.34)*||2.15 (0.58)||0.06 (0.14)*||0.02 (0.09)|
|Semantic||面包 – 蛋糕|
[Bread – Cake]
|2.19 (0.57)||4.04 (0.52)*||0.03 (0.07)||0.08 (0.15)*|
|Unrelated||司机 – 蛋糕|
[Driver – Cake]
|1.47 (0.49)||1.23 (0.21)||0 (0)||0 (0)|
|Nonword|| – 蛋糕|
[Kmbol – Cake]
The rating scores marked with asterisk in one condition is significantly higher than other conditions (Ps < 0.01). The number inside the parenthesis indicates standard deviation of the mean.
To ensure that stimulus materials met the above requirements, 4 normative assessments were collected: integration rating, similarity rating, free-combination, and free-association generation assessments. Sixty young adults who did not participate in the fMRI experiment provided responses to the first 2 ratings. The participants were asked to rate both the degree of integration and semantic similarity for all 504 word pairs (nonword condition was excluded). For the similarity rating, the participants were asked to judge how similar the words were in their perceptual or functional properties (from 1 for very different to 5 for very similar). For the integration rating, the participants were asked to judge how well the first word and the second word could be linked together to form a meaningful phrase (from 1 for totally unable to be linked to 5 for tightly linked). In addition, another 10 participants who did not participate in the fMRI experiment were recruited for another 2 stimulus assessments. In the free-combination generation task, the participants were required to write down a two-character noun that could be easily combined with a given noun to form a meaningful phrase (e.g., were shown Cherry and wrote Cake). In the free-association generation task, the participants were asked to write down a two-character noun that was semantically associated with a given noun but was not integrative (e.g., were shown Bread and wrote Cake). Finally, we calculated the proportion of the appearing word that is the same as the target word in all self-generated responses for both free-combination and free-association generation tasks.
Table 1 shows the statistical results of these 4 stimulus assessments. The word pairs in the integrative condition were more amenable to integration and had a higher free-combination proportion than in the unrelated (integration rating: t(167) = 80.7, P < 0.01; free-combination: t(167) = 6.8, P < 0.01) and semantic condition (integration rating: t(167) = 59.1, P < 0.01; free-combination: t(167) = 3.8, P < 0.01). In contrast, word pairs were both more similar and more frequently associated in the semantic condition than in the unrelated (similarity rating: t(167) = 74.2, P < 0.01; free-association: t(167) = 7.4, P < 0.01) and integrative condition (similarity rating: t(167) = 34.9, P < 0.01; free-association: t(167) = 4.4, P < 0.01). The condition specified for integration or similarity properties was further confirmed by directly comparing the integration ratings with the similarity ratings in the same condition. Specifically, there were significantly higher integration ratings than similarity ratings in the integrative condition (t(167) = 55.3, P < 0.01), while the reverse pattern was observed in the semantic condition (t(167) = 38.1, P < 0.01). Moreover, we strictly controlled and matched low-level linguistic variables of prime words across conditions: word frequency (mean ± SD, integrative = 36.99 ± 79.70; semantic = 36.63 ± 115.88; unrelated = 35.20 ± 76.90 per million; F2,168 = 0.10, P = 0.89), and total number of strokes (integrative = 15.84 ± 4.76; semantic = 15.94 ± 4.79; unrelated = 15.84 ± 4.49; F2,168 = 0.03, P = 0.97).
To maximize the integrative and semantic-priming effects and to avoid interference from preceding priming trials (Sachs et al. 2011), experimental trials were divided into 2 runs with 1 run devoted to the integrative stimuli and respective controls (integrative, unrelated, and nonwords), and the other to the semantic stimuli and respective controls (semantic, unrelated, and nonwords).
For the LDT, we added 112 word pair fillers where the target words were two-character pseudowords. To prevent participants from adopting a strategy during the experiment (e.g., they might tend to judge a target word as a real word when they read a nonword prime first), nonwords were used as the prime in half of the filler trials. Therefore, the number of filler trials with nonword primes was the same as the nonword trials.
To counterbalance stimuli across conditions and participants, 8 lists of the stimulus materials were constructed. No target word appeared twice in the same list, but the same target appeared in all conditions across participants. Given this scheme, the differences between conditions are not attributable to lexical properties of the targets. During the fMRI experiment, participants were required to complete one list of stimuli which consisted of 2 runs. The order of the 2 runs was also counterbalanced across participants to avoid any confounding from the presentation order. Therefore, each list contained a total of 280 trials (168 experimental trials and 112 filler trials).
Stimulus presentation and data collection was controlled by E-Prime (Psychology Software Tools, Pittsburgh, PA, USA; version 2.0). The stimuli were presented on a screen within the MRI cabin by a MRI-compatible LCD projector. The stimulus presentation schema is showed in the Figure 1A. Each trial constituted a white-color central fixation with a black background for 400 ms, a blank screen for 200 ms, a white-color prime word for 400 ms, a blank screen for 200 ms, and finally a yellow-color target word for 1500 ms followed by a 300-ms blank screen. This resulted in a stimulus-onset-asynchrony (SOA) of 600 ms. The participants were required to judge whether the yellow-color target words were real words or not by pressing a “yes” or “no” button with either their right index or middle ﬁnger, counterbalanced across participants. The target words disappeared once the participant made a response. In addition, to better estimate the hemodynamic response related to the onset of target words, we used jittered intertrial intervals, of 1, 3, or 5 s (average 3 s). Therefore, each trial lasted 6 s on average. We recorded the participants' response and reaction time (RT) in each trial during the fMRI experiment.
Functional Localizers for Predefining Language System
To predefine brain regions associated with language processing, we conducted 2 functional localization experiments before the LDT experiment. One was the word judgment task and another was a sentence reading task. The word judgment task was used to localize brain regions associated with lexical-semantic processing (Badre et al. 2005; Visser et al. 2012). The sentence reading task was used to localize brain regions associated with sentence-level processes, in which both semantic integrative and syntactic processing would be implicated. Details of these 2 functional localizers were descripted in the Supplementary Material.
MRI Data Acquisition
MRI data were acquired using a Siemens Trio 3T MRI system with a 32-channel head coil at the Shenzhen Institutes of Advanced Technology, Chinese Academic of Science. Three modalities of imaging data were collected, including functional, structural, and DTI images. To minimize signal loss and distortion in bilateral anterior temporal lobe (ATL) regions due to the magnetic susceptibility artifact, both low echo time (TE) (20 ms) and coronal slice orientation scanning parameters were applied (Axelrod and Yovel 2013) for the functional imaging runs. Specifically, the functional imaging were recorded by a T2*-weighted gradient echo-planar imaging (EPI) pulse sequence [repetition time (TR) = 2000 ms, TE = 20 ms, flip angle = 90°, 38 slices, field of view = 224 mm × 224 mm, in-plane resolution = 3.5 × 3.5, slice thickness = 3.5 mm with 1.1 mm gap]. T1-weighted high-resolution structural images were acquired using a magnetization-prepared rapid acquisition gradient echo sequence (176 slices, TR = 1900 ms, TE = 2.53 ms, ﬂip angle = 9°, voxel size = 1 × 1 × 1 mm3). Finally, DTI data with 30 diffusion encoding directions (TR = 10.8 s, TE = 87 ms, ﬂip angle = 90°, b = 1000 s/mm2) and one image without diffusion weighting (b value = 0 s/mm2, b0) were acquired. Each volume consisted of 85 slices in the intercommissural plane, 2-mm thickness with 2-mm gap, with an in-plane resolution of 2 mm and ﬁeld of view = 256 mm × 256 mm.
Functional MRI Data Analysis
All functional imaging data were preprocessed using SPM8 (Wellcome Department of Imaging Neuroscience, London, UK; www.fil.ion.ucl.ac.uk/spm/). The preprocessing procedure included slice-time correction, head-movement correction, coregistration between EPI and structural images, normalization to a standard T1 template in the Montreal Neurological Institute (MNI) space (resampling into 2 × 2 × 2 mm3 voxel size) and smoothing with a Gaussian kernel of 8-mm full width at half maximum.
We performed the subject-level analysis by using general linear modeling (GLM). Design matrices of the 3 tasks (2 localizers' tasks and the LDT) were constructed and modeled separately. In the 2 localizers, regressors of interest corresponding to word and nonword as well as sentence and nounlist trials were convolved with canonical hemodynamic response function (HRF) to build GLMs for lexical-semantic and the sentence integration effect, respectively. In the design matrix for the LDT task, 6 regressors of interest were included: integrative, unrelated, and nonword trials from the integrative run, and semantic, unrelated, and nonword trials from the semantic run. Both the filler trials and the incorrect response trials in each run were also modeled as noninterest regressors. The hemodynamic response at target onset was modeled for each of the 8 event types with the canonical HRF. In addition, low-frequency drifts were removed using a temporal high-pass filter (cutoff at 128 s). Six head-movement parameters were included in all design matrices as nuisance regressors to regress out motion-related artifacts. The standard gray matter volume created from the segmentation for each subject was used an inclusive mask to restrict voxels of interest.
A random-effect model was used for the group-level analyses. For functional localizers, we used a one-sample t-test to define brain regions associated with word-level semantic processes (word judgment task–nonword matching task) and sentence-level integration processes (sentence–nounlist) separately. Voxel-level P < 0.001 and cluster-level corrected P < 0.05 using family-wise correction were applied for multiple comparison correction. The union of these 2 contrasts was defined as a language mask. This mask was further used as an inclusive mask in the LDT task.
To calculate semantic and integrative effects, we constructed within-subject one-way Analysis of variance (ANOVA) to define brain regions showing a main effect either (or both) in the semantic and integrative runs in the LDT. The language mask created from the localizers was used as an inclusive mask in these 2 group-level ANOVAs to increase the statistical power. Monte Carlo simulations with the AlphaSim program were used to determine the activation threshold, taking into account both the number of voxel within the language mask and the smoothness of the preprocessed data (voxel-level P < 0.005, corrected to P < 0.05 with a 30-voxels cluster size, Cox 1996). The activation coordinates were reported in MNI space.
Regions of Interest Analyses and Parametric Modulation Analyses
To further investigate the priming effect in regions showing significant main effects in the ANOVAs, we conducted regions of interest (ROIs) analyses. Separate spherical ROIs with a 6-mm radius were created based on the peak coordinates of the brain regions revealed by the main effects in the ANOVAs. Such ROI construction approach is widely used (Poldrack 2007), and it can be used to ensure each ROI has the same number of voxel. The beta estimates of all the 6 conditions in the LDT were then extracted for each participant within each ROI using MarsBaR toolbox. The main aims of the ROI analyses were to detect which regions displayed significant priming effect (integrative vs. unrelated or semantic vs. unrelated) in each run and whether they showed dissociated priming effect (the contrast of the 2 types of priming effects).
In addition, we examined whether the regions associated with the integration or semantic effects were also related to the degree of integration or semantic association strength of word pairs by performing parametric modulation analyses. Two subject-level design matrices were constructed. The integration rating (or the semantic association) score for each trial from the stimulus assessments was used as a parametric modulation weight in the design matrix. The nonword trials, ﬁller trials, error response trials, and motion parameters were modeled separately as nuisance regressors. The parametric modulation analysis was performed at the subject-level first. Subsequently, we used the regions showing significant integrative or semantic effect as ROIs to extract beta estimates for each participant and conducted a one-sample t-test at group level. Moreover, we also conducted voxel-wise parametric modulation analyses to confirm the ROI results. Thus, the result would reveal which regions would show a monotonic modulation in activity as a function of integration or association strength.
DTI Data Analysis and Probabilistic Tractography
Diffusion-weighted imaging (DWI) data were analyzed using the FSL toolbox (Behrens et al. 2007) and functions from the AFNI package (Cox 1996). The DWI data were preprocessed for eddy currents and head motion using an affine registration model. Subsequently, the nonbrain tissues were removed using FSL's automated brain extraction tool (BET). In the tractography analysis, ROIs were selected from the regions showing significant semantic or integrative effects in the LDT experiment, and translated into white matter template. Here, we were interested in how the tempo-parietal regions anatomically connect to the frontal regions. Thus, a multiple-ROI approach was used (e.g., from ATL to anterior inferior frontal gyrus [aIFG]). The algorithm implemented in FSL (BedpostX) was first used to calculate the diffusion parameters for each voxel. After that, probabilistic tracking was performed by repeating 5000 random samples from the first ROI voxels to the second ROI voxels. These streamline samples started at the first ROI voxels and propagated through the local probability density functions of the estimated diffusion parameters. When a 2-ROI approach was used, only those streamlines initiated from the first ROI that reach a voxel in the second ROI (or vice versa) are retained.
Because we focused on within-hemisphere left fronto-temporal fiber connections, the right hemisphere was used as an exclusive mask, which has an effect of rejecting streamlines from the right hemisphere. All voxels within the left hemisphere will have a value representing the connectivity value between the first ROI voxels and the second ROI voxels (i.e., the number of samples that pass through that voxel). Probability maps were then normalized to the total number of ﬁbers (Upadhyay et al. 2007). Therefore, the probability maps represent the ﬁber density in the bundle between the 2 ROIs and are an indication of the most likely pathway between two gray matter regions.
Behavioral Performance in LDT
A one-way repeated-measures ANOVA was used for testing the condition differences in both accuracy and RT, with planned comparison threshold set at P = 0.05 after Bonferroni correction (Fig. 1B). No significant main effect (integrative, unrelated, and nonword) of accuracy in the integrative run was observed (F2,25 = 1.74, P = 0.19). In contrast, a marginally significant main effect of accuracy in the semantic run (semantic, unrelated, and nonword) was found (F2,25 = 2.91, P = 0.06). Further, post hoc comparisons found that the participants made less errors in the semantic trials than in both the unrelated (t(25) = 2.48, P = 0.02) and nonword trials (t(25) = 2.11, P = 0.05). For the RT analysis, both incorrect response trials and trials larger than 2.5 SD of the mean were removed. In the integrative run, a significant main effect of RT was found (F2,25 = 6.44, P = 0.003). Similarly, a marginally significant main effect of RT in the semantic run was observed (F2,25 = 2.93, P = 0.06). A post hoc planned comparison showed that the participants responded significantly faster in the integrative than in the unrelated condition, indicating an integrative-priming effect (26 ms; t(25) = 2.38, P = 0.025). Similarly, faster response time was also found in the pseudoword condition compared with the unrelated condition (t(25) = 3.18, P = 0.004). In the semantic run, participants responded significantly faster in the semantic condition than in the unrelated condition (20 ms; t(25) = 2.91, P = 0.007), indicating a semantic-priming effect. There was no significant difference between the unrelated and nonword condition (t(25) = 0.49, P = 0.62). Finally, we did not observe a significant difference between the integrative-priming and the semantic-priming effect (Fig. 1C; t(25) = 0.58, P = 0.56).
Brain Activations of the Functional Localizers
To ensure a high signal quality of the functional images, especially in the bilateral ATL regions, we calculated the temporal signal-to-noise ratio (tSNR, the ratio of the average signal intensity to the signal standard deviation across time points) for each voxel within the brain (Murphy et al. 2007). The results revealed that there was a good tSNR in the bilateral ATL (see Supplementary Fig. 1) such that most of the ATL regions were significantly higher than 40 (a minimal tSNR required for detecting condition differences).
Figure 2 presents the results of the 2 localizers. Figure 2A shows the distributed network activated during the word judgment task compared with the nonword matching task. These regions included left inferior frontal gyrus, a small portion of left anterior superior temporal gyrus (aSTG), left posterior middle temporal gyrus (pMTG), left tempo-parietal junctions (TPJs), posterior cingulate gyrus, precuneus, and right pMTG (see Fig. 2A, left panels: word > nonword). In addition, the comparison between the sentence and nounlist condition showed more distributed activations, including the left superior and middle frontal gyrus, bilateral inferior frontal gyrus, bilateral ATL, left pMTG, left TPJ, and dorsal medial superior frontal gyrus (see Fig. 2A, right panels: sentence > nounlist). Finally, a language mask was created by unifying these 2 localization maps (i.e., all regions activated in both contrasts were included in the mask, see Fig. 2B).
Functional Dissociation of Integrative and Semantic Priming Effects
Figure 3 illustrates the activation distributions in both the main effect of integrative (Fig. 3A) and semantic runs (Fig. 3B) within the language mask. The 2 brain maps were not thresholded to enable a visual comparison between the integrative and semantic effects. Figure 4A shows significant activations after applying a multiple comparisons correction in the 2 main effects. Eight regions revealed a significant semantic or integrative effect, all in the left hemisphere. Five regions showed a significant integrative effect, including aIFG, aSTG, ATL, pMTG, and TPJs. In contrast, another 3 regions showed a significant semantic effect, including posterior inferior frontal gyrus (pIFG), middle portion of middle temporal gyrus (mMTG), and posterior superior temporal gyrus (pSTG).
We performed a priming (semantic vs. integrative) by condition (priming, unrelated vs. nonword) ANOVA for each ROI first, followed by planed comparisons. These ANOVA analysis results showed that only the aIFG (F2,48 = 6.92; P = 0.009), aSTG (F2,48 = 10.87; P = 0.001), ATL (F2,48 = 9.48; P = 0.003), mMTG (F2,48 = 4.14; P = 0.043), and pSTG (F2,48 = 6.56; P = 0.012) showed significant priming-by-condition interactions.
|Regions||BA||MNI||Peak F-value||Number of voxels|
|Regions||BA||MNI||Peak F-value||Number of voxels|
aIFG, anterior inferior frontal gyrus; pIFG, posterior inferior frontal gyrus; aSTG, anterior superior temporal gyrus; pSTG, posterior superior temporal gyrus; mMTG, middle part of middle temporal gyrus; pMTG, posterior middle temporal gyrus; ATL, anterior temporal lobe; TPJ, temporal–parietal junctions. All these regions were located in left hemisphere.
To further examine whether those 8 regions showed a significant “dissociated priming effect” (i.e., more activation associated with one type of priming effect versus the other), we plotted the parameter estimates for each of the conditions (Fig. 4B,C) and further performed paired t-tests to directly compare the 2 types of priming effect (i.e., integrative–unrelated vs. semantic–unrelated, or unrelated–integrative vs. unrelated–semantic). We found 3 sets of regions. First, we found that only the aIFG (t(25) = 2.78, corrected P = 0.02), aSTG (t(25) = 3.16, corrected P = 0.008), and ATL (t(25) = 3.48, corrected P = 0.003) showed a significant dissociated integrative-priming effect. There was more activation (response enhancement) in the same 3 regions for the integrative compared with the unrelated condition.
Second, both the pSTG (t(25) = 2.55, corrected P = 0.05) and mMTG (t(25) = 2.78, corrected P = 0.04) showed a significant dissociated semantic-priming effect. The semantically related trials induced increased activity compared with the unrelated trials in these 2 regions as well. Furthermore, to test whether the semantic priming with response suppression in the aSTG was dissociated from its response enhancement in the integrative run, another contrast was performed: (unrelated–semantic)–(unrelated–integrative). The result showed a significant dissociated semantic-priming effect in the aSTG (t(25) = 2.95, corrected P = 0.01).
In addition, we further investigated whether the size of each behavioral priming effect was correlated with the level of each fMRI suppression or enhancement in these priming regions across subjects. We treated the behavioral priming effect (unrelated minus integrative or semantically related trials for each subject) as an independent variable and the fMRI priming as a dependent variable to build a regression model. The results showed that the magnitude of each behavioral priming effect was associated with the level of each fMRI priming for those priming regions (Supplementary Table 1). Specifically, the fMRI priming effect in both the aIFG and ATL were exclusively associated with the behavioral integrative-priming effect, while both the pSTG and mMTG were exclusively associated with semantic-priming effect.
In a third set of regions, we observed a main effect of relatedness (integrative or semantically related, unrelated and nonword) in the left pIFG (F2,25 = 4.49, corrected P = 0.03), pMTG (F2,25 = 4.43, corrected P = 0.04), and TPJ (F2,25 = 9.52, corrected P < 0.001) independent of the type of relationship between primes and targets. Both the pIFG and pMTG showed similar decreased response in the nonword condition compared with the related conditions (pIFG: t(25) = 2.99, corrected P = 0.005; pMTG: t(25) = 2.93, corrected P = 0.006) or unrelated conditions (pIFG: t(25) = 2.37, corrected P = 0.03; pMTG: t(25) = 2.53, corrected P = 0.02). However, there was no significant difference between the related and unrelated condition (pIFG: t(25) = 0.30, corrected P = 0.76; pMTG: t(25) = 0.01, corrected P = 0.88). In contrast, the TPJ showed different response patterns compared with the pIFG and pMTG. Significant increased response in the related condition compared with the unrelated one (t(25) = 4.34, corrected P < 0.001) was found. Similarly, increased response was found in the nonword condition compared with the unrelated condition (t(25) = 2.85, corrected P = 0.02).
Parametric Modulation of Integration Ratings and Semantic Association Strength
To test whether activity of those regions were also modulated by integration strength or semantic association, we extracted the beta estimates from the first-level parametric modulation analysis. The results showed that only the left ATL (t(25) = 4.29, corrected P = 0.001), TPJ (t(25) = 3.95, corrected P = 0.002), aSTG (t(25) = 3.51, corrected P = 0.008), and aIFG (t(25) = 2.94, corrected P = 0.03) showed a significant modulation effect of integration strength. That is, increasing integration rating was associated with increased activation in these regions. In contrast, the increased activation of the left pSTG (t(25) = 2.47, corrected P = 0.06), mMTG (t(25) = 2.67, corrected P = 0.04), and TPJ (t(25) = 2.60, corrected P = 0.05) showed significant modulation as a function of increased semantic association strength. There were no regions showing decreases of activation as a function of increases in either the integration ratings or semantic association strength. Voxel-wise parametric modulation analysis within the language mask further confirmed these ROI results (Supplementary Fig. 2).
To further verify whether the free-combination ratings could also account for the observed brain activations, we performed another parametric modulation analysis by using the free-combination ratings as the only regressor in the GLM. We observed that only the TPJ showed significant parametric modulation effect (t(25) = 3.18, corrected P = 0.015). In addition, we conducted another parametric modulation analysis using integrative ratings as regressor of interest while controlling for the free-combination ratings (noninterest regressor). We could still observe the 3 integrative-priming regions showing significant parametric modulation effect on integrative strength (aIFG: t(25) = 2.96, P = 0.026; aSTG: t(25) = 3.21, P = 0.015; ATL: t(25) = 4.30, P = 0.0009; TPJ: t(25) = 3.92, P = 0.002; pMTG: t(25) = 1.09, P = 0.71; pIFG: t(25) = 1.71, P = 0.34; mMTG: t(25) = 1.74, P = 0.32; pSTG: t(25) = 0.85, P = 0.83; All P-value were corrected for multiple comparisons).
Figure 5 summarizes all the results from probabilistic tractography, where mean normalized connectivity was thresholded at 3%, corresponding to the 95th percentile of the observed distribution (Wong et al. 2011). The tractography results showed that the brain regions associated with the integrative effect exhibited distinct fiber connection patterns from the pattern in the regions associated with the semantic effect (Fig. 5A,B). Specifically, the integrative-priming regions in temporal lobe (aSTG and ATL) connected to the aIFG through UF via the EC (Fig. 5A). In contrast, the semantic-priming regions in temporal lobe (mMTG and pSTG) connected to the pIFG by both the ECFS and the SLF/AF (Fig. 5B). Figure 5C summarizes 2 distinct fiber bundles associated with the integrative-priming and semantic-priming effects, respectively.
We found increased activation in the left aSTG, aIFG, and ATL when 2 concepts can be integrated into a unitary representation relative to those that cannot. This suggests that these regions contribute to integrative priming. These integrative-priming regions were different from the regions associated with semantic priming in both their localization and response patterns. Tractography analyses further indicated that fiber connections within the integrative-priming regions and those within the semantic-priming regions were separated (Fig. 6). These findings offer the first empirical evidence in healthy humans that both different regions and potential brain circuits support integrative versus semantic priming.
Although lexical-semantic priming effects have been consistently observed behaviorally, the neural mechanisms of this phenomenon are still unclear. The primary reason is that it is not straightforward to associate the neural (or BOLD) activity to the behavioral priming. Theoretically, similar behavioral priming might be a consequence of different neural processes or associated with different neural mechanisms (Henson 2003; Hutchison 2003; Estes and Jones 2009). Here, we used custom fMRI scanning parameters to overcome signal distortion in anterior temporal cortices, and demonstrated such possibility by separating the neural substrates associated with the 2 behavioral priming effects, and further uncovered that the integrative strength and semantic association might be underlying factors driven both behavioral and neural priming effects. These findings are inline with the theoretical predictions and also provide us an opportunity to understand the neural formation (functional segregation and anatomical interaction across regions) of the semantic system, and how these circuits associated with the representation and integration of semantic information.
Anterior Fronto-Temporal Regions Associated with Integrative-Priming Effect
RTs were faster in the integrative condition compared with the unrelated condition. The effect size of this behavioral integrative priming is comparable with previous findings (Estes and Jones 2009; Mather et al. 2014), suggesting that we have successfully manipulated the prime–target relationship and replicated the integrative-priming effect. Enhanced activity in the integrative condition relative to the unrelated condition was found in the left aSTG, aIFG, and ATL. The involvement of these 3 regions has been associated with semantic integrative processes in previous studies on sentence comprehension. For example, they have been frequently observed activated when contrasting sentences versus lexical-level baseline (Humphries et al. 2006; Rogalsky and Hickok 2009; Pallier et al. 2011) and incongruent sentences versus congruent ones (Tesink et al. 2009; Zhu et al. 2012, 2013). In our localization experiment, we also replicated the findings using the contrast of sentences versus nounlists (Fig. 2).
The neural response enhancement in these regions might be associated with the relational integration processes that could facilitate the lexical decision on the target words. Such response enhancement has been proposed to be associated with additional processes linked to the formation of new representations (Henson 2003; Sachs et al. 2011; Lee et al. 2014). Indeed, cognitive processes of integrative word pairs have been assumed to involve additional components compared with the unrelated pairs. According to the relational integration hypothesis (Estes and Jones 2009; Mather et al. 2014), 2 critical components, the complementary role activation and combinatorial processing, are involved during integrative priming (Estes and Jones 2009; Mather et al. 2014). In line with this hypothesis, we found similar enhanced response patterns in both left aSTG and aIFG, which indicate they are both associated with integrative processing. In addition, studies on local phrase composition align with our interpretation, showing that the aSTG is involved in the process of building local phrase structures (Friederici et al. 2000, 2003; Grodzinsky and Friederici 2006; Friederici 2011). Similarly, the semantic integration role of the left IFG (particularly its anterior portion locating in the BA 47), are supported by previous findings on sentence comprehension, in which the aIFG was activated when participants were required to integrate semantic information from different information sources (e.g., speaker identities and world knowledge) (Hagoort et al. 2004; Tesink et al. 2009). Furthermore, activation in left aIFG has been parametrically modulated by semantic integration load in sentence comprehension, which was independent of task manipulations (Zhu et al. 2012) and general executive control processes (Zhu et al. 2013). Here, we replicated these findings and further revealed activities of both aSTG and aIFG increasing as a function of increasing integration strength between primes and targets. Altogether, these results suggested that both the left aSTG and aIFG play important roles in integrating semantic information to create a unitary representation.
It worth to note that previous studies have found stronger activation in the left aIFG for higher semantic integration load (Hagoort et al. 2004; Tesink et al. 2009; Zhu et al. 2012, 2013), while here we found increased activation with increased prime–target integration strength. This seemingly opposite effect may reflect how the integration component is involved in different task contexts. Here, we used a priming paradigm with relatively short SOA, in which participants were not required to explicitly detect prime–target relationships, or to try to integrate word pairs when they encountered unrelated pairs. Thus, the integration component is less likely implicated for the unrelated than the integrative pairs. In contrast, most of previous studies used sentence comprehension task and explicitly asked participants to judge the congruency of sentences. In this task context, the integration load could be increased for the incongruent conditions that were similar with the unrelated condition in our study. Therefore, the integration component is more likely implicated for the incongruent than the congruent sentence.
In contrast, left ATL showed both a robust dissociated integrative-priming effect and a lexical-semantic effect (increased activation in the nonword condition compared with the unrelated condition), which were different from that found in the left aSTG and aIFG. Consistent with our findings, ATL activation was frequently observed during sentence comprehension (Humphries et al. 2006; Rogalsky and Hickok 2009; Pallier et al. 2011). Lesions in these regions are associated with difficulty in sentence-level language comprehension (Dronkers et al. 2004). In addition, Bemis et al. (2011, 2013) have found increased activation in the ATL during the two-word composition compared with the one-word composition task in early time windows (150–250 ms) by using magnetoencephalography (MEG). The evidence together with our findings suggests that the left ATL activation is associated with basic elementary integrative processing.
Another possible explanation is that the ATL may encode integrative relationships stored in long-term memory. It has been proposed that ATL plays a role as a core semantic hub (Rogers et al. 2004; Patterson et al. 2007; Jefferies 2013). This semantic hub would function as a convergence zone for distributed features of concepts, responsible for integrating feature information from modality-specific regions (e.g., regions response specifically to colors, shapes and movements, etc.) (Patterson et al. 2007; Pereira et al. 2009; Correia et al. 2014; Coutanche and Thompson-Schill 2015; Lambon Ralph 2014). This “convergent representation” hypothesis implies that ATL might play a role in representing semantic relationships between concepts and their features. Indeed, recently researchers have found that multivoxel patterns in the ATL encode semantic relationships between features (such as “green” and “round”) and object identity (such as “lime”) (feature-to-identity links) (Clarke and Tyler 2014; Coutanche and Thompson-Schill 2015). Similarly, the integrative-priming effect observed in ATL might rely on the same mechanism. In the integrative condition, the prime word acts as one of a feature property of the target word (Cherry – Cake) while the target word acts as the main concept. Retrieval of this “feature-to-identity” relational information resulted in increased activation in the integrative word pairs relative to the unrelated condition. Future studies are required to disentangle the relationship between the basis semantic integrative processing and the semantic representation in the ATL (Westerlund and Pylkkänen 2014).
Posterior Temporal Regions Associated With Semantic-Priming Effect
We also found the classic semantic-priming effect in terms of both reduced error rate and facilitated RT in the semantically related word pairs relative to the unrelated ones. Importantly, we also found both neural response suppression and enhancement in the semantic condition relative to the unrelated condition in temporal cortex.
Semantic priming has been associated with spreading activation of semantic information across concepts via conceptual similarity or association relationship (Lucas 2000; Hutchison 2003). In our study, the left aSTG showed neural response suppression (i.e., unrelated > semantic), in accordance with the pattern of behavioral facilitation. In addition, such neural response suppression in the aSTG is also consistent with previous findings using the masked priming paradigm (Lau et al. 2013; Ulrich et al. 2013) and semantic priming with short SOA (Rissman et al. 2003; Sass, Krach, et al. 2009; see review in Lau et al. 2008). Researchers have proposed that response suppression during semantic priming might be associated with spreading activation via feature overlap across concepts. When the prime and target share semantic features on functional or perceptual dimensions, activation of these features in the primes could ease the activation of the target concepts, resulting in neural response decreases (Henson 2003; Sachs et al. 2011).
In contrast, neural response enhancement was observed in both the left pSTG and mMTG (i.e., semantic > unrelated), suggesting they might play distinct roles during semantic priming compared with the aSTG. There is increasing evidence that semantic priming not only leads to response suppression but also to response enhancement (Kotz et al. 2002; Raposo et al. 2006; Sass, Krach, et al. 2009; Sass, Sachs, et al. 2009; Sachs et al. 2011; Lee et al. 2014), especially in the left temporal regions observed here (Kotz et al. 2002; Raposo et al. 2006; Sass, Krach, et al. 2009). Such response enhancement has been attributed to different or additional processes linked to the formation of new associations or representations (Henson 2003). Given that these two temporal regions have been also associated with strategic semantic-priming processes (Badre et al. 2005; Gold et al. 2006), such as semantic relationship (association or similarity) detection (Badre et al. 2005; Raposo et al. 2006), we interpret such response enhancements in both mMTG and pSTG as related to association-based semantic activation. Consistent with this interpretation, we observed that parametrically increased association strength between primes and targets was related to enhanced activation in these two regions. Here, such association-based activation is considered to reflect controlled component of semantic-priming processing but we did not exclude the possibility of involving automatic spreading activation for these regions. Automatic priming mainly occurs at short stimulus-onset asynchronies (SOAs) and have been associated with the automatic spreading activation across semantic memory (Collins and Loftus 1975). In contrast, controlled semantic priming (at a long SOA) have been considered to reflect the postlexical strategic semantic processes (Gold et al. 2006). Nevertheless, the automatic spreading activation can still occur at a long SOA, but the controlled processes will dominate. Altogether, different response patterns in these temporal regions suggest a functional segregation in the semantic processing system (Binder et al. 2009; Price 2012).
Common Effect During the Semantic and Integrative-Priming
In contrast to the dissociated priming effects, there was a main effect of condition in the left pIFG, pMTG and TPJ independent of the type of prime–target relationships. Both the pIFG and pMTG showed similar response patterns, in which we found decreased activations in the nonword condition relative to the other two conditions. While the semantic-priming effect with both neural enhancement (Sachs et al. 2011; Whitney et al. 2011; Lau et al. 2013; Lee et al. 2014) and suppression (Gold et al. 2006; Liu et al. 2010) has been observed, these two regions have also been associated with controlled semantic processes (Badre et al. 2005; Gold et al. 2006; Ye and Zhou 2009; Whitney et al. 2011). Specifically, increased activations have been found when more competitors are involved. Indeed, it has been suggested that the left pIFG contributes to maintaining or inhibiting irrelevant internal representations while the pMTG has been associated with controlled semantic retrieval (Whitney et al. 2011; Zhu et al. 2013). Such an explanation was further confirmed in the present study. Our results showed that the two regions are sensitive to the manipulation of the number of presented words (2 words > 1 word condition) rather than conceptual relationships. Therefore, more conceptual meanings would be retrieved and maintained in the two-word conditions, whereby further selection processes might be involve to manipulate the representation based on the task goals. Taken together, these findings suggest pIFG and pMTG might be related to a common postlexical controlled semantic process during integrative and semantic priming.
The left TPJ was frequently found to be activated in varieties of semantic tasks such as lexical-semantic and sentence comprehension (Binder et al. 2009; Price 2012). We replicated such findings in our two localizers. In addition, previous studies have found that this region is sensitive to the manipulation of constituting individual meanings (Pallier et al. 2011) and building semantic association (Seghier et al. 2010). Consistent with these observation, we found that the activation of left TPJ was sensitive to both the number of words and amount of relatedness (related > unrelated), suggesting that the left TPJ may play a general role in maintaining semantic information and be involved in semantic associations regardless of an integrative or semantic relationship.
Separated Brain Circuits Associated with the Two Types of Priming Effects
One novelty of our results is that we showed possible pathways associated with integrative and semantic priming. The regions associated with the two types of priming effects were not only partially separated in localization, but also showed different brain circuitry patterns.
Probabilistic fiber tracking showed that the left temporal regions associated with dissociated integrative-priming effect (ATL, aSTG, and aIFG) were connected with each other via only the UF running through the EC. Such fiber connection was consistent with observations from both neurotypical adults and semantic dementia (SD) patients. In neurotypical adults, the temporal pole regions, especially Brodmann Area (BA) 38, connects to frontal operculum (BA 47) via UF (Friederici et al. 2006; Thiebaut de Schotten et al. 2012). Moreover, SD patients not only suffered from ATL atrophy but also showed decreased white matter integrity of the UF that connected the ATL and the anterior frontal regions (Agosta et al. 2009). Furthermore, most of the previous DTI studies did not precisely define the functions of the activated region used for fiber tractography (but see Griffiths et al. 2012 for defining syntactic processing regions). For example, Saur et al. (2008) only used normal sentences compared with meaningless sentences to define regions of interest that were associated with semantic processing. In contrast, here we precisely isolated regions associated with integrative priming and characterized the functional roles of these regions based on their neural response patterns. Our findings further extend the knowledge concerning the UF pathway where it connects three regions (ATL–aSTG–aIFG) associated with integrative priming.
In contrast, pSTG, mMTG, and aSTG connect to pIFG via two language-related fiber bundles, ventral (ECFS) and dorsal pathways (SLF/AF). Consistent with this observation, previous studies have found that the language-related temporal cortices connect to the frontal cortices via both ventral and dorsal pathways (Saur et al. 2008; Rolheiser et al. 2011; Dick and Tremblay 2012; see review in Friederici 2012). These two fiber bundles are not limited to semantic priming but also support other language functions. For instance, the SLF/AF fiber bundle has been associated with speech repetition (Saur et al. 2008), complex syntactic processing (Wilson et al. 2011), lexical-semantic processing (Glasser and Rilling 2008), and reading ability (Zhang et al. 2014). Similarly, the ECFS has been associated with sentence-level semantic processing (Saur et al. 2008) and “sound-to-meaning” association learning (Wong et al. 2011). Here, it is possible that the left pSTG, mMTG, aSTG, and pIFG connect with each other via both the SLF/AF and ECFS forming an interconnected language circuit supporting spreading activation of semantic information.
This work was supported by grants from the Natural Science Foundation of China (31271086 and 31300834), and key project from the Natural Science Foundation of Guangdong Province, China.
We thank Shaowei Guan for her assistance with stimuli preparation and fMRI data collection. We also thank Erica Middleton and Tom Verguts for their constructive comments and helpful language editing of our earlier version of the manuscript. Conflict of Interest: None declared.