www.elsevier.com/locate/ynimg NeuroImage 34 (2007) 1270 – 1279 Semantic ambiguity processing in sentence context: Evidence from event-related fMRI Monika-Zita Zempleni, a,b,c,⁎ Remco Renken, b,c John C.J. Hoeks, a,b,c Johannes M. Hoogduin, b,c and Laurie A. Stowe a,b,c a Department of General Linguistics, University of Groningen, The Netherlands School of Behavioral and Cognitive Neurosciences, University of Groningen, The Netherlands c Neuroimaging Center, University of Groningen, The Netherlands b Received 6 May 2005; revised 14 September 2006; accepted 21 September 2006 Available online 4 December 2006 Lexical semantic ambiguity is the phenomenon when a word has multiple meanings (e.g. ‘bank’). The aim of this event-related functional MRI study was to identify those brain areas, which are involved in contextually driven ambiguity resolution. Ambiguous words were selected which have a most frequent, dominant, and less frequent, subordinate meaning. These words were presented in two types of sentences: (1) a sentence congruent with the dominant interpretation and (2) a sentence congruent with the subordinate interpretation. Sentences without ambiguous words served as a control condition. The ambiguous words always occurred early in the sentences and were biased towards one particular meaning by the final word(s) of the sentence; the event at the end of the sentences was modeled. The results indicate that a bilaterally distributed network supports semantic ambiguity comprehension: left (BA 45/44) and right (BA 47) inferior frontal gyri and left (BA 20/37) and right inferior/ middle temporal gyri (BA 20). The pattern of activation is most consistent with a scenario in which initially a frequency-based probabilistic choice is made between the alternative meanings, and the meaning is updated when this interpretation does not fit into the final disambiguating context. The neural pattern is consistent with the results of other neuroimaging experiments which manipulated various aspects of integrative and context processing task demands. The presence of a bilateral network is also in line with the lesion and divided visual field literature, but contrary to earlier claims, the two hemispheres appear to play similar roles during semantic ambiguity resolution. © 2006 Elsevier Inc. All rights reserved. Keywords: Functional MRI; Left hemisphere; Right hemisphere; Frontal lobe; Temporal lobe; Language; Lexical semantic ambiguity; Homograph; Disambiguation; Context; Integration ⁎ Corresponding author. Institute of Neuroradiology, University Hospital of Zurich, Fraueklinikstrasse 10, 8091 Zurich, Switzerland. E-mail address: monika.zempleni@usz.ch (M.-Z. Zempleni). Available online on ScienceDirect (www.sciencedirect.com). 1053-8119/$ - see front matter © 2006 Elsevier Inc. All rights reserved. doi:10.1016/j.neuroimage.2006.09.048 Introduction In natural language, many words have multiple meanings, for instance, the English word ‘bank’ means a monetary institute as well as ground next to a river. This phenomenon is referred to as lexical semantic ambiguity. Words with alternative meanings which are spelled identically are also called homographs. Although native speakers encounter ambiguous words very frequently, they usually do not perceive them as ambiguous, unless ambiguity is emphasized as in puns for example. The lack of detection is presumably due to fast and automatic disambiguation processes, i.e. ambiguity resolution, under which one particular meaning of the ambiguous word is chosen. The literature suggests that there are two sources of information on which this choice is based. The first is the context into which the interpretation has to fit; the second is frequency-based meaning dominance, i.e. the relative frequency of the alternative meanings. Depending on whether one meaning is much more frequent than the alternatives or the alternative meanings are equally frequent, an ambiguous word can be unbalanced or balanced in frequency. Various models of ambiguity resolution differ in the importance they assign to context and frequency-based dominance during ambiguity resolution. (For reviews on models and theories, see e.g. Gorfein, 2001 or Coney and Evans, 2000.) Hereby we will focus on the exhaustive access model proposed by Onifer and Swinney (1981), which has received the most empirical support and which provided the theoretical framework for the paradigm of the current study. According to the exhaustive access model, all meanings become active initially, regardless of frequency-based dominance. This view is supported by numerous semantic priming studies (e.g. Swaab et al., 2003; Swinney and Love, 2001). Semantic priming refers to the phenomenon where a word meaning is more easily accessed if the meaning of a semantically related word, i.e. prime, has already been activated; this can be reflected in e.g. shorter response times. Exhaustive meaning access is evinced by the fact that reading an ambiguous word (e.g. ‘bank’) can facilitate comprehension of words related to both alternative meanings M.-Z. Zempleni et al. / NeuroImage 34 (2007) 1270–1279 (e.g. ‘money’ and ‘river’) irrespective of context. That is, both context incongruent and context congruent meanings can cause priming effects (at least for a couple of hundreds of milliseconds; see studies below). Following exhaustive access of several meanings, frequency-based dominance and contextual congruity lead to a resolution to the appropriate meaning. The goal of the current experiment is to examine the neural bases of ambiguity resolution. Earlier studies suggest that the neural substrate of ambiguity resolution might be of particular interest for a neuroimaging experiment as both hemispheres are suggested to be involved in this cognitive process. Burgess and Simpson (1988) evaluated the effects of frequency-based meaning dominance and the time course of disambiguation in a divided visual field experiment. They presented homographs centrally without context as primes followed by a target word presented laterally; the targets thus were initially available only to one hemisphere. The results indicated that both meanings were able to facilitate the recognition of targets immediately (35 ms) in the left hemisphere, but only the dominant meaning continued to elicit priming at a longer delay (750 ms). In the right hemisphere, only the dominant meaning led to an immediate priming effect, but both meanings facilitated the targets at the long delay. The authors concluded that the left hemisphere quickly accesses both meanings and soon focuses on the dominant one, while in the right hemisphere, access might be slower, particularly for the subordinate meaning, and that the right hemisphere either does not choose between the alternatives or does so more slowly. Faust and Chiarello (1998) also used the divided visual field paradigm to investigate hemispheric differences in the use of context during ambiguity resolution. They found that the left hemisphere is able to select the contextually appropriate word meaning within a short period (900 ms), as opposed to the right hemisphere. The authors suggested that the right hemisphere may maintain both meanings, which might be useful when revision of the initially selected interpretation is necessary. The divided visual field technique supports hemispheric specialization, but gives no information about the brain regions which are involved. The few lesion studies available suggest that Broca’s area might be necessary for the quick access of multiple meanings and that damage to this area at least slows ambiguity resolution (Swaab et al., 1998; Swinney et al., 1989). Left nonthalamic subcortical and cortical (predominantly fronto-parietal and fronto-temporal) damage apparently disrupts context-driven processes during semantic ambiguity resolution (Copland et al., 2000, 2002). Damage to either anterior temporal lobe (Zaidel et al., 1995) or to various areas in the right hemisphere (e.g. basal ganglia, fronto–temporo–parietal region) (Grindrod and Baum, 2005) has also been shown to cause deficits in lexical semantic disambiguation. Relatively few neuroimaging experiments have yet addressed the neural substrates of ambiguity resolution. Copland et al. (2003) studied the neural substrate of semantic priming using f MRI with a test material in which the primes were ambiguous words. The target words were either related to the dominant or the subordinate meaning of the homograph. They compared the related conditions to an unrelated target condition. They found less activation, due to priming, in both related target conditions in the left middle temporal gyrus and in the left inferior prefrontal cortex, which suggests priming effects from both meanings, which provides neuroimaging support for the exhaustive access view. However, Copland et al. (2003) did not include an unambiguous condition to 1271 explicitly investigate the effect of ambiguity per se. Chan et al. (2004) added unambiguous words to their study and compared them to ambiguous words in a semantic relatedness generation task. Ambiguous words elicited increased activation predominantly in bilateral middle and superior frontal gyri and in the anterior cingulate, whereas unambiguous words elicited predominantly bilateral inferior frontal and temporal increases in activation. Ambiguity resolution in sentential context was studied by Rodd et al. (2005) using f MRI and by Stowe et al. (2005) using PET. Rodd et al. (2005) found left posterior inferior temporal cortex and bilateral inferior frontal gyri activation, whereas Stowe et al. (2005) found only right inferior frontal activation extending into right temporal pole during ambiguity resolution. The results of these neuroimaging studies only partially overlap; due to the different designs, however, it is likely that they addressed different aspects of ambiguity resolution. We will come back to these results in more detail in the Discussion section in comparison to the results of the current study. The main goal of the current study was to further clarify the neural substrates of homograph comprehension using event-related functional MRI. In the current paradigm, unbalanced homographs were presented in a sentential context. The ambiguous words always occurred at the beginning of sentences, and care was taken that the words preceding the homograph did not provide context biasing toward either meaning (i.e. neutral). The homograph was followed by several additional words of neutral context. The final word(s) of the sentence provided the biasing context, supporting either the most frequent, dominant, or a less frequent, subordinate interpretation of the homograph. Thus, there were sentences with a dominant meaning congruent final context, hence dominant sentences (D), and sentences with a subordinate meaning congruent context, hence subordinate sentences (S). Matched sentences without ambiguous words served as a control condition (C). See Table 1, D, S, C conditions. Therefore, the novelty of the current design compared to the earlier studies is that meaning dominance is explicitly manipulated. As said above, the theoretical framework in which the paradigm was created is the exhaustive access model (Onifer and Swinney, 1981), which proposes that multiple word meanings are accessed initially and kept active until a choice can be made between the alternatives. Although this model proposes that multiple meanings can be accessed even in the presence of preceding biasing context, we additionally facilitated exhaustive access in the current paradigm by the sentential structure: the homographs occurred Table 1 Examples of the three experimental conditions with literal English translations Dominant sentence (D) Subordinate sentence (S) Control sentence (C) De advocaat werd ∣ door het oudje ∣ keurig benaderd. The lawyer was/by the old-person/nicely approached. De advocaat werd ∣ door het oudje ∣ keurig ingeschonken. The egg liqueur was/by the old-person/nicely poured-out. Het gras werd ∣ door de tuinman ∣ nooit gemaaid. The grass was/by the gardener/never mowed. Note. The homograph is underlined; the biasing context is typeset in bold; vertical lines indicate presentational phrase borders. 1272 M.-Z. Zempleni et al. / NeuroImage 34 (2007) 1270–1279 early in the sentences, without any preceding biasing context. What we intended to test as the secondary aim of the current study was whether the meaning choice is exclusively driven by biasing context, which in this experiment occurs downstream toward the end of the sentence, or whether in the case of the unbalanced homographs used in this experiment an earlier choice can be carried out, based on frequency-based meaning dominance, in which case the context must ‘overwrite’ this initially favored meaning if necessary. The first scenario we considered is that, at the point of disambiguation, both meanings are equally available. If this is the case, the ambiguous conditions are expected to be more difficult than the unambiguous sentences, but the two ambiguous conditions are not expected to differ from each other because a choice has to be made between alternatives in both cases. In brain areas which support this process, we expect to see a pattern of activation in which S = D > C. Hence, this pattern will be interpreted as reflecting selection between equally available alternatives (selection). The second hypothesis considered here is that, even though the initial meaning access is exhaustive, still the alternative word meanings are not kept equally active for an unlimited time. Instead, an initial choice is made between the different interpretations on the basis of frequency-based meaning dominance. If this favored interpretation does not fit into the final disambiguating context, an alternative meaning must be reactivated and integrated to the context. If the homographs were 100% unbalanced, we might expect a pattern of activation in which S > D = C because the dominant meaning would always be chosen initially and the meaning update would only be necessary for the subordinate condition. However, the homographs used were not unbalanced to this degree (see Materials section below), so it is likely that some subjects initially favor the subordinate meaning in some sentences at least. The sum of these two processes would therefore lead to a pattern in which S > D > C. Thus, the S > D = C or S > D > C patterns are both consistent with a scenario of an initial frequency-based choice followed by updating the word meaning and reintegrating the alternative with the disambiguating context. These patterns will hence be referred to as the meaning update (update) patterns. Note that areas showing these relative patterns of conditions would only indicate where the meaning update takes place but cannot be interpreted as sites where the initial meaning choice was carried out. Based on the results of the divided visual field literature (see above), we were particularly interested in whether there is any evidence that context- and frequency-based meaning dominance are utilized differently in the two hemispheres; e.g. is there any evidence that alternative word meanings are more available in the right than in the left hemisphere? Materials and methods Materials The materials consisted of three types of Dutch sentences: sentences with a semantically ambiguous word (homograph) biased either toward its dominant or toward its subordinate meaning, and structurally identical sentences without ambiguous words serving as a control. Eighty homographs (fifty nouns and thirty verbs) were used in the ambiguous sentences. The dominance of the alternative meanings was identified using the results of an unpublished word meaning association study by Bruinsma et al. (manuscript).1 In this offline study, forty-six naive native speakers were asked to write down the first word that came to mind on reading each homograph. Then, the number of associates related to each alternative homograph meaning was determined. For example, if the Dutch word ‘advocaat’ (eggliqueur or lawyer) was most often associated with words referring to ‘courts’ or ‘law’ and less frequently referring to ‘alcohol’ and ‘drink’, then the meaning ‘lawyer’ was considered to be dominant and the meaning ‘egg-liqueur’ was considered to be subordinate. The difference in dominance was fairly strong; dominant-related associations were given in 80.3% of the cases. Two sentences were constructed containing each homograph as illustrated in Table 1. In both versions, the homograph was presented early in the sentence (underlined in the table), followed by a neutral, non-biasing context that lasted at least three words, and the sentence finished with a context, usually one word long (bold in table), which strongly biased the homograph either towards its dominant or subordinate interpretation. As can be seen in Table 1, the difference between the dominant and subordinate sentence versions, the biasing context, was usually only one word. To check that the non-biasing context is indeed neutral, we first consulted a panel of ten expert native speakers (linguists and linguistics students) who confirmed that disambiguation to both meanings remained possible after the intervening words. Then, a Cloze procedure test was carried out by twenty naive native speakers on the sentence contexts without the final disambiguating word(s). The results indicated that frequency-based meaning dominance was not substantially altered by the intervening words. On average, 72% of the sentence fragments were completed employing the dominant meaning (median percentage 82.5%), in 21.6% of the cases, the completions were consistent with the subordinate meaning (median percentage 12.5%), which is approximately the ratio indicated by the associate data. The remaining responses were either ambiguous between the two readings or were unclassifiable for other reasons. Last, the plausibility ratings for the two conditions were similar (see below), indicating that both versions of the final biasing context were considered appropriate at the end of the sentence fragments. If the intervening context created a bias toward one meaning, the other completion should come across as implausible. Unambiguous control sentences were constructed, each with the same syntactic structure as two of the homograph sentence pairs. Each sentence was broken into several phrases for presentation; matched sentences were divided into phrases identically (generally three phrases). Over all items, the dominant, subordinate, and control sentences were matched for sentence structure, sentence plausibility, sentence length, and average word frequency. The word frequencies were taken from the CELEX data base (Baayen et al., 1993). Plausibility was determined using an off-line rating list with a five-point scale (1 = totally implausible and 5 = totally plausible). The characteristics of the experimental conditions are summarized in Table 2. In order to avoid repetition effects, the sentences were assigned to two experimental lists, which contained either the dominant or the subordinate version of a homograph sentence pair; this is why twice as many homograph sentences were created as control sentences. The two lists contained the same set of control sentences (40); the dominant and subordinate sentence pairs (80) were 1 For further information on the Bruinsma et al. study, please contact the corresponding author. M.-Z. Zempleni et al. / NeuroImage 34 (2007) 1270–1279 distributed so that subjects saw only one version of a sentence pair (40 dominant and 40 subordinate sentences). The sentence conditions on the two lists were also matched for sentence structure, plausibility, phrase number, length in letters, and for average word frequency. Thirty-six filler sentences with various structures were added to both lists. Ambiguous words also occurred in some filler sentences. After each filler sentence, a target word was presented, as illustrated in Table 3, and subjects decided whether it was related to the sentence in meaning. Half of the targets were related to the fillers, whereas the other half were unrelated to them. The semantic relationship between the fillers and targets was intentionally rather easy, so a poor performance on this task would indicate lack of cooperation or attention. This task was also intended to distract subjects from the homograph sentences. As the task was imposed on the filler sentences only, the hemodynamic response function to the target stimuli should not be affected by motor responses or decision strategies. The order of the stimuli in each list was pseudo-random, with items in each condition spread evenly across the entire list and no more than two items of the same condition presented consecutively. In order to minimize effects of fatigue and lack of concentration, the experiment was divided into four equally long runs. The order of the four runs was counterbalanced over subjects resulting in 2 lists by 2 presentation orders. Apparatus and procedure The sentences were presented visually, phrase-by-phrase (see Tables 1 and 3), and centered on the computer screen. For programming and presentation, we used E-Prime (Psychology Software Tools Inc., 2001). The text was printed in black on a white background, with a font size which allowed for comfortable reading. A projector transmitted the stimuli from the computer to a screen which was visible to the subjects via a coil-mounted mirror. Subjects were instructed to read and comprehend the sentences silently. In order to monitor attention and cooperation during the scanning session, participants were also asked to carry out the relatedness decision described above (see filler sentences in Table 3). Responses were made by pressing buttons on a response box with the index and middle fingers of the right hand. Each item was presented as follows. First, a fixation cross was presented in the center of the screen, simultaneously with the beginning of a scan of a brain volume. This remained until replaced by the first phrase of the sentence; sentence onsets were jittered between 0 and 3 s, which was the volume acquisition (TR) time, to decouple the sequence of slices from the onset of stimuli. Each phrase remained centered on the screen for approximately 100 ms per character and then was replaced by the following phrase. After the final phrase, a fixation cross was displayed until the beginning of the following item, which was initiated by a trigger from the MR scanner. The filler sentences were followed by Table 2 Condition match of the test materials Condition Number of sentences Dominant (D) 80 Subordinate (S) 80 Control (C) 40 Plausibility Length in letters Average word frequency 4.14 (0.45) 3.83 (0.59) 4.39 (0.42) 55 (10) 55 (10) 55 (10) 2.27 (1.07) 2.30 (1.05) 2.07 (1.27) 1273 Table 3 Examples of the filler sentence—target word pairs with literal English translations Filler sentencerelated target Filler sentenceunrelated target Toen het donker werd ∣ besloten de soldaten ∣ te stoppen.—oorlog When it dark became, ∣ decided the soldiers ∣ to stop.—war Tijdens de bevalling ∣ moest de vrouw ∣ al haar energie gebruiken.—gordijn During the childbirth ∣ must the woman ∣ all her energy use.—curtain Note. Vertical lines indicate presentational borders. a 600 ms fixation cross then the target appeared for 3000 ms and then another fixation that lasted until the next item. The total presentation time for the sentences was 5.5 s on average (SD = 1.2 s, minimum = 3 s, maximum = 10 s) for each condition. The total time for each fixation–sentence–fixation combination was 21 s, the time necessary to acquire seven complete brain volumes. The long interstimulus interval was chosen to allow the hemodynamic response to return to baseline before initiating the following trial. The measurements were carried out using a Philips Intera 3 T MRI in the Neuroimaging Center of the University of Groningen. Echo planar images (T2* weighted) were acquired (TR = 3000 ms, TE = 35 ms, flip angle = 90°). Each volume consisted of 46 slices covering the whole brain (slice thickness = 3.5 mm, slice gap = 0 mm, field of view 224 × 224 × 161 mm, in plane resolution: 64 × 64). Homograph sentence comprehension was tested after the scanning session using a relatedness decision procedure, similar to that used in the scanner, but in this case imposed on the ambiguous sentences. A subset of ambiguous sentence pairs from the scanner materials were selected, and three target words were chosen for each of them, based on the associated words collected by Bruinsma et al. (manuscript).1 One target was related to the dominant but not to the subordinate meaning, another was related to the subordinate but not to the dominant meaning, and one target was unrelated to either meaning. This produced six conditions: the dominant sentence followed by the congruent dominant-related target (D-con); the dominant sentence followed by the incongruent subordinate-related target (D-inc); and the dominant sentence followed by the unrelated target (D-unr); and the same for the subordinate sentences, i.e. S-con, S-inc, and S-unr. For example, the sentence ‘De advocaat werd door het oudje keurig benaderd/ ingeschonken’ (‘The lawyer/egg liqueur was by the old person nicely approached/poured out’) was followed by the word ‘meubel’ (furniture) as unrelated target or the words ‘rechtbank’ (‘court’) versus ‘drankje’ (‘drink’) as congruent or incongruent targets depending on the dominant versus subordinate contexts.2 The behavioral material was programmed and presented the same way as the materials in the scanner (see above), and the task was implemented using a pc. Subjects were instructed to respond as accurately as possible by pressing the control and left arrow keys of the pc using the index and middle fingers of the right hand. The average error rate was calculated for each subject in order to establish whether the subject was capable of comprehending the test sentences. 2 For more information on the behavioral dataset, please contact the corresponding author. 1274 M.-Z. Zempleni et al. / NeuroImage 34 (2007) 1270–1279 The experiment was approved by the Medical Ethical Committee of the University Medical Center Groningen. Participants gave written informed consent prior to participating in accordance with the Helsinki Declaration. Before inclusion, the participants were screened for MRI incompatibility, and during scanning, the standard MRI safety regulations were followed. Subjects Data from sixteen subjects were included in the study, after excluding one subject from the seventeen scanned due to movement artifacts. All participants were healthy, native Dutch speakers (8 females, 8 males; average age 32.5 years, SD = 10.1; average educational level 15.8 years, SD = 4.1, handedness score 0.99, SD = 0.1, indicating almost exclusive right handedness (Van Strien, 1992)). All subjects had normal or corrected to normal vision, normal hearing, and no history of neurological or psychiatric disorder. Data analysis and model specification The raw f MRI data were converted into analyze format using the MRIcro software package (Rorden and Brett, 2000). SPM2 (Wellcome Department of Cognitive Neurology, London, UK) was used for spatial preprocessing, such as realignment, normalization (2 × 2 × 2 mm voxel size), and smoothing with a 10 mm Gaussian kernel (Friston, 1994). The realigned data were checked for movement artifacts; translation movements bigger than 5 mm and rotation movements bigger than 3° were rejected causing exclusion of one subject out of seventeen scanned. Since the aim of the study was to evaluate the neural substrate of contextual disambiguation, we modeled the point of disambiguation as an event. The biasing context becomes available at the end of the sentences (in 90% of the sentences, the disambiguating word is the final word). Therefore, the end of the presentation of the last phrase was used as the time point of this event. In 10% of the sentences, additional words followed the disambiguating one in order to make the sentences more plausible; the presentation time of these extra words varied between 300 ms and 1300 ms (760 ms on average). Considering the 3000 ms volume acquisition time, this short additional presentation time in a low percent of sentences is very unlikely to influence the results, particularly given individual variability in reading speed and reading strategies. SPM2 was used for statistical analysis, during which beta estimates were calculated for each subject for each condition (D, S and C) modeled as events at the sentence end-point (first level analysis). These beta estimates were entered into a random effects one-way ANOVA on the second level (Henson and Penny, 2005; Penny et al., 2003), during which a non-sphericity correction was applied. After identifying significant areas of hemodynamic response, a further analysis was carried out in order to identify which conditions contributed to the significant difference. Basically, we were interested whether the subordinate and dominant conditions differed significantly in any of the areas which were significant in the F-test since this would exclude the possibility of the selection hypothesis, which would suggest equal availability of the alternative meanings at the sentence end-points. For this reason, we created an inclusive mask of the F-test and the S > D t-test. Additionally, we checked if any of the areas showed significant difference between both the S > D and the D > C contrasts. For this reason, we made a conjunction analysis of these two contrasts. Results Behavioral performance in the scanner and on the homograph comprehension task f MRI participants performed well on the relatedness decision task in the scanner (imposed on filler sentences; see Apparatus and procedure section). The good performance indicates that participants were attentive and cooperative during scanning (average error rate: 0.06, SD: 0.06, minimum = 0, maximum = 0.017). Therefore, no dataset had to be left out due to inappropriate inscanner behavioral performance. Subjects also performed well on the post scanner homograph comprehension task (see Apparatus and procedure section) as indicated by their low error rates across all conditions (average = 0.12, SD = 0.06, minimum = 0.03, maximum = 0.23). This suggests that participants were able to comprehend the homograph sentences and to choose the context congruent meaning of the homographs. Therefore, no dataset had to be left out due to inappropriate behavioral performance. f MRI results Table 4 lists those brain areas which showed significant effects at p uncorrected ≤ 0.0002 (F ≥ 11.47) and consisted of at least 5 voxels. The areas are organized according to anatomical regions. kE refers to cluster size expressed as number of voxels. F stands for the maximum F value within the cluster. Lat stands for hemispheric lateralization; LH: left hemisphere; RH: right hemisphere. Montreal Neurological Institute coordinates (x, y, z) and the Brodmann areas (BA) are provided in separate columns. As can be seen, five clusters showed significant hemodynamic change at this threshold: two clusters in the left inferior frontal gyrus, in fact two maxima in a single larger area as indicated in Fig. 1 (see below); one in the left inferior/middle temporal gyrus, in the right inferior/ middle temporal gyrus, and in the right inferior frontal gyrus. The areas are also shown in Fig. 1, projected on a standard anatomical template (MRIcro; Rorden and Brett, 2000). The picture shows the areas at a slightly lower threshold (see color bar) in order to show that the two activation maxima in the left inferior Table 4 Anatomical regions showing significant effects during homograph comprehension kE F Lat Anatomical region Frontal lobe 14 12.38 LH Inferior frontal gyrus 9 12.38 LH Inferior frontal gyrus 6 12.88 RH Inferior frontal gyrus x − 48 − 52 34 y z 26 20 16 26 20 −10 BA 45 44 47 Temporal lobe 31 15.50 LH Inferior/middle temporal gyri − 50 − 48 −12 20/37 18 13.21 RH Inferior/middle temporal gyri 56 − 34 −16 20 Note. The analysis was time-locked to the end of each sentence, to the point of disambiguation. All areas were significant at p ≤ 0.0002 (F ≥ 11.47), the spatial extent of activation (kE) was ≥5 voxels. Areas are presented with Montreal Neurological Institute coordinates (x, y, z), the cytoarchitectural designation according to Brodmann (BA), the maximum F value (F), of the hemodynamic response in the area and its extent (kE). In order to emphasize laterality effects, the hemispheric lateralization (Lat) is presented in a separate column. LH: left hemisphere; RH: right hemisphere. M.-Z. Zempleni et al. / NeuroImage 34 (2007) 1270–1279 1275 Fig. 1. Notes: The picture above shows the anatomical regions whose activation maximum showed significant hemodynamic response during disambiguation at a threshold of p uncorrected ≤ 0.0002 (F ≥ 11.47) and kE ≥5. The picture shows the areas at a slightly lower threshold (see color bar) in order to show that the two activation maxima in the left inferior frontal gyrus (area IV) are in fact part of one continuous region. The statistical map is projected on a standard anatomical template (MRIcro; Rorden and Brett, 2000). BA: Brodmann area. The plots show the zero-mean corrected magnitude of the effect of conditions, averaged over the whole cluster in each area. The error bars show the 90% confidence interval. D: dominant condition; S: subordinate condition; C: unambiguous control condition. frontal gyrus (area IV on the figure) are in fact two maxima of a continuous area. This figure also shows the contrast estimates in the different regions. As already can be seen on the plots, the significance is mainly due to the effect of the subordinate sentence condition. In order to determine the different patterns, we created an inclusive mask of the F-test and the S > D contrast. This revealed that the subordinate condition elicited significantly higher (p uncorrected ≤ 0.0002; F ≥ 11.47; kE ≥ 5) activation than the dominant condition in each of the areas. Additionally, the areas in the left inferior frontal gyrus showed significance when evaluated in a conjunction analysis of the S > D and the D > C contrasts at a lower threshold (p uncorrected ≤ 0.005; T ≥ 2.75; and kE ≥ 5). These results suggest that the relative pattern of conditions is most consistent with an S > D > C pattern in the left inferior frontal gyrus and with the S > D = C pattern in the other areas. There is no indication for the S = D > C pattern in any of the areas. Discussion In this f MRI experiment, the neural substrate of lexical semantic ambiguity comprehension was studied, with a special focus on the hemispheric contribution to context-driven disambiguation. The results indicate that a bilaterally distributed network underlies context-driven semantic ambiguity resolution. The majority of the activated voxels are located in the left hemisphere, in the left inferior frontal gyrus (BA 45/44) and in the left inferior/middle temporal gyri 1276 M.-Z. Zempleni et al. / NeuroImage 34 (2007) 1270–1279 (BA 20/37). Significant activations were also seen, however, in the right hemisphere, in the right inferior frontal gyrus (BA 47), and in the right inferior/middle temporal (BA 20) gyri. These findings are consistent with the results of earlier neuroimaging experiments which evaluated ambiguity comprehension in sentential context, in particular, with the findings of Rodd et al. (2005). In an f MRI study, they compared sentences containing several ambiguous words (with disambiguating information in the sentence) versus sentences without ambiguous words. The high ambiguity versus unambiguous sentence contrast elicited activation in the left posterior inferior temporal gyrus and in the left inferior frontal gyrus; the left hemisphere activations were very similar to those we found in the current experiment. They also reported right inferior frontal activation but in a more lateral and superior region than in the current study. Stowe et al. (2005) presented ambiguous words in the beginning of sentences and disambiguated them at the end of the sentences to the subordinate meaning. They compared this condition to an unambiguous sentence condition in a blocked design PET experiment and found right inferior frontal activation; this area was also significant in the current study. The lack of a more extensive activation in the Stowe et al. (2005) study might be due to the slower time course of PET which makes it less sensitive to relatively short responses within the block or to the fact that each ambiguous sentence was disambiguated to the subordinate meaning making the task more transparent for the participants. To sum up, three of the areas identified in the current study (left and right inferior frontal gyri and the left inferior/middle temporal gyri) were also activated in other studies, which studied semantic ambiguity resolution in sentence context, although the experimental designs and methodology were not identical. The occurrence of bilateral activation in both the frontal and the temporal lobes is likely due to the novel element of our design; in the current paradigm, frequency-based meaning dominance and contextual disambiguation were systematically manipulated which might have increased the processing demand recruiting additional areas. In contrast to the studies above, another neuroimaging study on semantic ambiguity comprehension by Chan et al. (2004) yielded a completely different set of areas. They presented blocks of semantically ambiguous and unambiguous words, without context, to which participants had to covertly generate semantically related words. Ambiguous words elicited increased activation predominantly in bilateral middle and superior frontal gyri and in the anterior cingulate, whereas unambiguous words elicited predominantly bilateral inferior frontal and temporal activation. However, in this study, the experimenters did not provide disambiguating context, which would force participants to choose a particular meaning that is congruent with this context. In this case, therefore, subjects could only base their decision when generating a related word on the relative frequency of the alternative meanings. Taken together, these data suggest that the presence of any disambiguating information may profoundly influence the neural network involved in comprehension of semantically ambiguous words. In other words, disambiguation on purely probabilistic grounds with no need for contextual integration presents a different cognitive scenario calling for a different neural network than the situation when one particular meaning has to be chosen according to the context. Therefore, the network which we see here appears to be responsible for contextual updating/integration. This hypothesis is clearly supported by the S > D = C and the S > D > C patterns seen in the activated areas. As pointed out in the Introduction, these patterns suggest that the network is primarily concerned with updating the sentence meaning when context turns out to be inconsistent with an initially favored meaning. A support for this interpretation of our data is that the areas in the current study have also been found to be significant in other neuroimaging studies which have manipulated other aspects of semantic/contextual integrative demand, for example, by introducing anomalies into the test sentences or using unexpected sentence completions. Kuperberg et al. (2003) found that the left inferior frontal gyrus (BA 44/45, 47, 9/46) is sensitive to pragmatic violations whereas semantic violations elicited right middle and superior temporal gyri activation (BA 21/22) in another study (Kuperberg et al., 2000). These activations are similar to those found in the current study: left BA 45/44 and a slightly more ventral region in the right temporal lobe (right inferior/middle temporal gyrus; BA 20). In another f MRI study, Baumgartner et al. (2002) manipulated the degree to which sentence final nouns were expected as sentence completions (expected, unexpected, and violated sentence completions).The unexpected and anomalous conditions elicited activations in the left inferior frontal and left posterior middle temporal gyri, in areas similar to those we found in the current study. It has to be noted that the sentences in our study were all judged to be plausible in an off-line rating, but the subordinate sentences might have temporarily been detected as implausible and in any case they had less expected sentence completions by definition. Additionally, the left middle temporal gyrus has also been reported to be involved in the comprehension of complex and semantically less transparent phrases and sentences (Homae et al., 2002). Last, bilateral frontal and temporal areas, with left dominance, have been implicated in generating the event-related potential occurring in the 300–500 ms range (N400) which is extremely sensitive to factors which affect semantic integration difficulty such as expectancy violations or increased demand for semantic integration. See for example Halgren et al., 2002 or Van Petten and Luka, 2006 for recent reviews. Turning to the functions of the individual areas, the involvement of the prefrontal cortex during context-based semantic ambiguity resolution is not surprising given that this brain area is generally implicated in context processing (Cohen and ServanSchreiber, 1992). Additionally, the left inferior frontal gyrus, mostly BA 44 and 45 areas, has been suggested to subserve a general mechanism for selecting among competing semantic representations; its involvement in selection has been demonstrated using different cognitive tasks, e.g. verbal fluency (Hirshorn and Thompson-Schill, 2006), verb generation, object classification and semantic comparison (Thompson-Schill, 2003). This area is also involved in dealing with competing syntactic representations as can be seen in activations elicited by syntactic garden-path sentences (Mason et al., 2003; Stowe et al., 2004). Badre et al. (2005) demonstrated that the left mid-ventrolateral prefrontal cortex (BA 45) is crucial in mnemonic processing and supports a general post-retrieval selection mechanism, i.e. selecting relevant knowledge from competing information, which is a distinct cognitive process from the top–down retrieval of semantic knowledge localized more to the left anterior ventrolateral prefrontal cortex (BA 47). From a wider perspective, selection between competing representations may call on some aspect of working memory. The involvement of the left inferior frontal gyrus in the current M.-Z. Zempleni et al. / NeuroImage 34 (2007) 1270–1279 paradigm is in line with the (verbal) working memory literature (D’Esposito et al., 2000). There is considerable empirical evidence suggesting that the left ventrolateral prefrontal cortex, including Broca’s area, is crucial for maintaining and manipulating information in working memory (for maintenance, see Fiez et al., 1996; Paulesu et al., 1993; for manipulation, see Gelfand and Bookheimer, 2003). Left ventrolateral prefrontal cortex also appears to support executive control processes in working memory, such as overcoming proactive interference, i.e. inhibiting prepotent response tendencies (D’Esposito et al., 2000; Jonides et al., 1998). In the current homograph comprehension paradigm, the dominant word meaning can be considered as a prepotent response; therefore, the involvement of the BA 45 area may reflect the necessity for actively suppressing the dominant meaning and allowing information update and the reintegration of the less potent subordinate word meaning; reintegration and updating may also make use of verbal maintenance processes.3 With regard to the right inferior frontal lobe, its role in linguistic context updating is more tendentious, but recent neuroimaging studies have shown that, when task demands increase during language processing, the right inferior frontal gyrus tends to get involved, e.g. degree of speech compression (Poldrack et al., 2001) or when subjects were asked not only to comprehend but also to ‘repair’ syntactic violation (Meyer et al., 2000). Note that this latter study involved a process which is in some sense similar to ‘repairing’ an initial ‘misinterpretation’ during semantic ambiguity resolution. Although context processing in general has been associated with prefrontal regions, language related semantic integration which requires context updating has also been associated with temporal regions. For example, many studies, which applied sentential or lexical semantic manipulations, have found left inferior/middle temporal activations (see for example the Kuperberg and the Baumgartner studies cited above; or for a review, see Stowe et al., 2005). Rodd et al. (2005) also found a similar posterior area during their semantic ambiguity comprehension task (see above). Involvement of the right temporal lobe has not been reported as frequently, but several studies suggest that, under certain circumstances, it responds to increased context processing demands, e.g. Kuperberg et al. (2000), see above. In another f MRI study, Kircher et al. (2001) asked subjects to complete sentence context without a missing final word; the sentence stems were characterized by low Cloze probability. In the generation condition, subjects had to generate a plausible sentence ending. In the decision condition, subjects had to choose between two words, presented by the experimenters. The control condition was reading complete sentences. The generation condition versus the other two conditions contrast elicited, among other areas, extensive right lateral temporal activation. The decision versus reading contrast revealed left inferior frontal and bilateral middle–superior temporal gyri activations at a slightly more dorsal location than in our study. The authors argue that multiple meanings have to be accessed and evaluated during this task, especially in the generation condition, because the sentence stems had low Cloze 3 Working memory has typically been investigated using delayed response tasks. These tasks are hypothesized to elicit several cognitive subcomponents, such as information encoding, maintenance, manipulation, retrieval, response selection and inhibition. There is evidence that these subprocesses are carried out by different parts of the lateral prefrontal cortex (D’Esposito et al., 2000). 1277 probability. Since our material requires the access and evaluation of alternative interpretations, our findings are in line with this suggestion. Finally, let us consider the bilateral nature of the neural network involved. This finding is consistent with the lesion literature and explains why both left hemisphere damage (Copland et al., 2000, 2002; Swaab et al., 1998; Swinney et al., 1989) and right hemisphere damage (Grindrod and Baum, 2005; Zaidel et al., 1995) can disrupt lexical semantic ambiguity comprehension. Similarly, this bilateral network involvement is in line with the proposals based on divided visual field experiments (Faust and Chiarello, 1998; Burgess and Simpson, 1988). However, the relative condition patterns do not provide support for the claims that the two hemispheres contribute in a substantially different manner. Our results did not support the hypothesis that, at least in the right hemisphere, the alternative meanings are equally available for use during context updating because the S = D > C pattern is not seen in any of the significant areas. It has to be noted, however, that the studies mentioned above evaluated the availability of the alternative meanings at shorter delays (900 ms in the former and 750 ms in the latter study) whereas in the current study the presentation time of the intervening phrase(s) which contained a neutral context was 1835 ms on average. Taken together, this evidence may indicate that the right hemisphere is capable of carrying out contextual disambiguation, but it does so more slowly than the left hemisphere. An alternative explanation that is more consistent with the model proposed by Faust and Chiarello (1998) is that the left hemisphere carries out the initial choice and the right hemisphere only participates in the reactivation of the alternative meanings. Further studies are needed, applying for example combined f MRI and EEG, to more precisely determine the time course of context-driven disambiguation in the two hemispheres. The results reported here suggest that frequency-based meaning dominance and contextual congruency invoke two totally different processes in the brain. However, in order to substantiate this view, the role of these two factors has to be investigated more systematically in future studies. For example, by varying the degree of imbalance between the availability of the different meanings and the degree to which context supports these meanings, we could get a clearer picture of the areas which respond to these factors. Since these factors are graded rather than discrete, treating them as variables within a regression analysis conducted over items could identify areas which parametrically vary in intensity of activation due to these variables. Conclusions The present study investigated the neural substrate of lexical semantic ambiguity in sentence context. The results indicate the involvement of a bilaterally distributed network: bilateral inferior frontal gyri and bilateral inferior/middle temporal gyri. These findings are consistent with the lesion literature as well as with existing neuroimaging data. The pattern of activation suggests that, even if multiple meanings are accessed during the comprehension of ambiguous words, they are not maintained for an unlimited time. Instead, a meaning frequency-based choice is most likely, at least in case of unbalanced homographs. If a later biasing context occurs which does not match the initially favored choice, the alternative word meaning has to be updated and integrated to the context. 1278 M.-Z. Zempleni et al. / NeuroImage 34 (2007) 1270–1279 Acknowledgments We would like to thank Marco Haverkort for contributing to the test materials; members of the Center for Language and Cognition Groningen group for valuable comments on an earlier draft. Remco Renken is a Hazewinkel-Beringer fellow; Monika-Zita Zempleni has been supported by an Ubbo Emmius scholarship provided by the Faculty of Arts, University of Groningen. References Baayen, H.R., Piepenbrock, R., Van Rijn, H., 1993. The CELEX Lexical Database. Linguistic Data Consortium, Philadelphia, PA. Badre, D., Poldrack, R.A., Pare-Blagoev, E.J., Insler, R.Z., Wagner, A.D., 2005. Dissociable controlled retrieval and generalized selection mechanisms in ventrolateral prefrontal cortex. Neuron 47, 907–918. Baumgartner, A., Weiller, C., Buchel, C., 2002. Event-related f MRI reveals cortical sites involved in contextual sentence integration. NeuroImage 16 (3), 736–745. Burgess, C., Simpson, G.B., 1988. Cerebral hemispheric mechanisms in the retrieval of ambiguous word meanings. Brain Lang. 33 (1), 86–103. Chan, A.H., Liu, H.L., Yip, V., Fox, P.T., Gao, J.H., Tan, L.H., 2004. Neural systems for word meaning modulated by semantic ambiguity. NeuroImage 22, 1128–1133. Cohen, J.D., Servan-Schreiber, D., 1992. Context, cortex, and dopamine: a connectionist approach to behavior and biology in schizophrenia. Psychol. Rev. 99 (1), 45–77. Coney, J., Evans, K.D., 2000. Hemispheric asymmetries in the resolution of lexical ambiguity. Neuropsychologia 38 (3), 272–282. Copland, D.A., Chenery, H.J., Murdoch, B.E., 2000. Processing lexical ambiguities in word triplets: evidence of lexical–semantic deficits following dominant nonthalamic subcortical lesions. Neuropsychology 14 (3), 379–390. Copland, D.A., Chenery, H.J., Murdoch, B.E., 2002. Hemispheric contributions to lexical ambiguity resolution: evidence from individuals with complex language impairment following left-hemisphere lesions. Brain Lang. 81, 131–143. Copland, D.A., de Zubicaray, G.I., McMahon, K., Wilson, S.J., Eastburn, M., Chenery, H.J., 2003. Brain activity during automatic semantic priming revealed by event-related functional magnetic resonance imaging. NeuroImage 20, 302–310. D’Esposito, M., Postle, B.R., Rypma, B., 2000. Prefrontal cortical contributions to working memory: evidence from event-related f MRI studies. Exp. Brain Res. 133, 3–11. Faust, M., Chiarello, C., 1998. Sentence context and lexical ambiguity resolution by the two hemispheres. Neuropsychologia 36 (9), 827–835. Fiez, J.A., Raife, E.A., Balota, D.A., Schwarz, J.P., Raichle, M.E., 1996. A positron emission tomography study of the short-term maintenance of verbal information. J. Neurosci. 16, 808–822. Friston, K.J., 1994. Statistical parametric mapping. In: Thatcher, R.W., Hallett, M. (Eds.), Functional Neuroimaging: Technical Foundations. Academic Press, Inc., pp. 79–93. Gelfand, J.R., Bookheimer, S.Y., 2003. Dissociating neural mechanisms of temporal sequencing and processing phonemes. Neuron 38 (5), 831–842. Gorfein, D.S., 2001. On the Consequences of Meaning Selection: Perspectives on Resolving Lexical Ambiguity. American Psychological Association. Grindrod, C.M., Baum, S.R., 2005. Hemispheric contributions to lexical ambiguity resolution in a discourse context: evidence from individuals with unilateral left and right hemisphere lesions. Brain Cogn. 57, 70–83. Halgren, E., Dhond, R.P., Christensen, N., Van Petten, C., Marinkovic, K., Lewine, J.D., Dale, A.M., 2002. NeuroImage 17, 1101–1116. Henson, R., Penny, W., ANOVAs and SPM. 2005. http://www.fil.ion.ucl.ac. uk/∼wpenny/publications/rik_anova.pdf. Hirshorn, E.A., Thompson-Schill, S.L., 2006. Role of the left inferior frontal gyrus in covert word retrieval: neural correlates of switching during verbal fluency. Neuropsychologia 44, 2547–2557. Homae, F., Hashimoto, R., Nakajima, K., Miyashita, Y., Sakai, K.L., 2002. From perception to sentence comprehension: the convergence of auditory and visual information of language in the left inferior frontal cortex. NeuroImage 16 (4), 883–900. Jonides, J., Smith, E.E., Marshuetz, C., Koeppe, R.A., Reuter-Lorenz, P.A., 1998. Inhibition in verbal working memory revealed by brain activation. Proc. Natl. Acad. Sci. U. S. A. 95, 8410–8413. Kircher, T.T.J., Brammer, M., Andreu, N.T., Williams, S.C.R., McGuire, P.K., 2001. Engagement of right temporal cortex during processing linguistic context. Neuropsychologia 39, 798–809. Kuperberg, G.R., McGuire, P.K., Bullmore, E.T., Brammer, M.J., RabeHesketh, S., Wright, I.C., Lythgoe, D.J., Williams, S.C.R., David, A.S., 2000. Common and distinct neural substrate for pragmatic, semantic, and syntactic processing of spoken sentences: an f MRI study. J. Cogn. Neurosci. 12 (2), 321–341. Kuperberg, G.R., Holcomb, P.J., Sitnikova, T., Greve, D., Dale, A.M., Caplan, D., 2003. Distinct patterns of neural modulation during the processing of conceptual and syntactic anomalies. J. Cogn. Neurosci. 15 (2), 272–293. Mason, R.A., Just, M.A., Keller, T.A., Carpenter, P.A., 2003. Ambiguity in the brain: what brain imaging reveals about the processing of syntactically ambiguous sentences. J. Exper. Psychol. Learn. Mem., Cogn. 29 (6), 1319–1339. Meyer, M., Friederici, A.D., von Cramon, D.Y., 2000. Neurocognition of auditory sentence comprehension: event related f MRI reveals sensitivity to syntactic violations and task demands. Cogn. Brain Res. 9, 19–33. Onifer, W., Swinney, D.A., 1981. Accessing lexical ambiguities during sentence comprehension: effects of frequency of meaning and contextual bias. Mem. Cogn. 9 (3), 225–236. Paulesu, E., Frith, C.D., Frackowiak, R.S.J., 1993. The neural components of the verbal component of working memory. Nature 362, 342–344. Penny, W., Holmes, A., Friston, K., 2003. Random effects analysis, In: Frackowiak, R., Friston, K., Frith, C., Dolan, R., Price, C., Zeki, S., Ashburner, J., Penny, W. (Eds.), Human Brain Function, 2nd ed. Academic Press. Poldrack, R.A., Temple, E., Protopapas, A., Nagarajan, S., Tallal, P., Merzenich, M., Gabrieli, J.D.E., 2001. Relations between the neural bases of dynamic auditory processing and phonological processing: evidence from f MRI. J. Cogn. Neurosci. 13 (5), 687–697. Rodd, J.M., Davis, M.H., Johnsrude, I.S., 2005. The neural mechanisms of speech comprehension: f MRI studies of semantic ambiguity. Cereb. Cortex 15 (8), 1261–1269. Rorden, C., Brett, M., 2000. Stereotaxic display of brain lesions. Behav. Neurol. 12, 191–200. Stowe, L.A., Paans, A.M.J., Wijers, A.A., Zwarts, F., 2004. Activations of “motor” and other non-language structures during sentence comprehension. Brain Lang. 89 (2), 290–299. Stowe, L.A., Haverkort, M., Zwarts, F., 2005. Rethinking the neural basis of language. Lingua 115 (7), 997–1042. Swaab, T., Brown, C., Hagoort, P., 1998. Understanding ambiguous words in sentence contexts: electrophysiological evidence for delayed contextual selection in Broca’s aphasia. Neuropsychologia 36 (8), 737–761. Swaab, T., Brown, C., Hagoort, P., 2003. Understanding words in sentence contexts: the time course of ambiguity resolution. Brain Lang. 86 (2), 326–343. Swinney, D., Love, T., 2001. Context effects on lexical processing during auditory sentence comprehension: on the time-course and neurological bases of a basic comprehension process. In: Witruk, W., Friederici, A.D., Lachmann, T. (Eds.), Basic Functions of Language, Reading and Reading Disabilities. Kluwer Academic Publishers, Dordrecht. Swinney, D., Zurif, E., Nicol, J., 1989. The effects of focal brain damage on sentence processing: an examination of the neurological organization of a mental module. J. Cogn. Neurosci. 1, 25–37. M.-Z. Zempleni et al. / NeuroImage 34 (2007) 1270–1279 Thompson-Schill, S.L., 2003. Neuroimaging studies of semantic memory: inferring “how” from “where”. Neuropsychologia 41 (3), 280–292. Van Petten, C., Luka, B.J., 2006. Neural localization of semantic context effects in electromagnetic and hemodynamic studies. Brain Lang. 97 (3), 279–293. 1279 Van Strien, J.W., 1992. Classificatie van links-en rechtshandige proefpersonen. Nederlands Tijdschrift voor de Psychologie 47, 88–92. Zaidel, D.W., Zaidel, E., Oxbury, S.M., Oxbury, J.M., 1995. The interpretation of sentence ambiguity in patients with unilateral focal brain surgery. Brain Lang. 51 (3), 458–468.