Skip to main page content
Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
, 50 (4), 499-513

Lexical and Syntactic Representations in the Brain: An fMRI Investigation With Multi-Voxel Pattern Analyses

Affiliations

Lexical and Syntactic Representations in the Brain: An fMRI Investigation With Multi-Voxel Pattern Analyses

Evelina Fedorenko et al. Neuropsychologia.

Abstract

Work in theoretical linguistics and psycholinguistics suggests that human linguistic knowledge forms a continuum between individual lexical items and abstract syntactic representations, with most linguistic representations falling between the two extremes and taking the form of lexical items stored together with the syntactic/semantic contexts in which they frequently occur. Neuroimaging evidence further suggests that no brain region is selectively sensitive to only lexical information or only syntactic information. Instead, all the key brain regions that support high-level linguistic processing have been implicated in both lexical and syntactic processing, suggesting that our linguistic knowledge is plausibly represented in a distributed fashion in these brain regions. Given this distributed nature of linguistic representations, multi-voxel pattern analyses (MVPAs) can help uncover important functional properties of the language system. In the current study we use MVPAs to ask two questions: (1) Do language brain regions differ in how robustly they represent lexical vs. syntactic information? and (2) Do any of the language bran regions distinguish between "pure" lexical information (lists of words) and "pure" abstract syntactic information (jabberwocky sentences) in the pattern of activity? We show that lexical information is represented more robustly than syntactic information across many language regions (with no language region showing the opposite pattern), as evidenced by a better discrimination between conditions that differ along the lexical dimension (sentences vs. jabberwocky, and word lists vs. nonword lists) than between conditions that differ along the syntactic dimension (sentences vs. word lists, and jabberwocky vs. nonword lists). This result suggests that lexical information may play a more critical role than syntax in the representation of linguistic meaning. We also show that several language regions reliably discriminate between "pure" lexical information and "pure" abstract syntactic information in their patterns of neural activity.

Figures

Figure 1
Figure 1
Four experimental conditions that have been used extensively in previous neuroimaging studies to study lexical and syntactic processes. Sample items are taken from Fedorenko et al. (2010; Experiment 1).
Figure 2
Figure 2
Idealized functional profiles of response for a voxel or a region that is sensitive to both lexical and syntactic information (purple frame), to lexical information only (red frame), and to syntactic information only (blue frame). S=sentences, W=Word-lists, J=Jabberwocky sentences, N=Nonword lists.
Figure 3
Figure 3
Five sample scenarios for how a functional profile for a region that is sensitive to both lexical and syntactic information (e.g., purple frame in Fig. 2) can arise. Each 4×4 grid represents a brain region, and each square in the grid represents a voxel. Red voxels are sensitive to lexical information only, blue voxels are sensitive to syntactic information only, and purple voxels are sensitive to both lexical and syntactic information.
Figure 4
Figure 4
(adapted from a figure by Julie Golomb): A schematic illustration of the logic of the correlation-style multi-voxel pattern analyses (Haxby et al., 2001).
Figure 5
Figure 5
Top: A probabilistic overlap map showing in each voxel how many of the 25 individual subjects show a significant (at p p< .05, FDR-corrected) effect for the Sentences p> Nonwords contrast. Bottom: The main functional parcels derived from the probabilistic overlap map using an image parcellation (watershed) algorithm, as described in more detail in Fedorenko et al. (2010).
Figure 6
Figure 6
Left: The results of the searchlight analyses showing discriminability between Sentences and Nonword-lists conditions (black frame), and pairs of conditions that differ along the lexical dimension (S vs. J and W vs. N; red frame) or along the syntactic dimension (S vs. W and J vs. N; blue frame). [Note of course that the contrasts that involve the Sentences condition – S vs. N, S vs. J and S vs. W – additionally involve compositional semantic processes, which may be contributing to discriminability.] Right: The group-level (random-effects) activation maps for the corresponding contrasts. All the maps are thresholded at p<.05, FDR-corrected (i.e., the darker red colors show voxels that reach this significance level, with the brighter colors showing voxels that reach higher significance levels), except for the Jabberwocky> Nonwords maps which are thresholded at p<.001, uncorrected, because no voxels emerged – either for the searchlight- or activation-based analysis – for this contrast at the FDR .05 threshold.
Figure 7
Figure 7
Differences between within- and between-condition correlation values for pairs of conditions that differ along the lexical dimension (red bars) and pairs of conditions that differ along the syntactic dimension (blue bars). For each ROI and each subject we averaged the within- vs. between- difference scores for pairs of conditions that differ along the lexical dimension (i.e., S-SJ, W-WN, J-JS, N-NW) and for pairs of conditions that differ along the syntactic dimension (i.e., S-SW, W-WS, J-JN, N-NJ). We then averaged these values across subjects for each ROI. Error bars represent standard errors of the mean over subjects. Asterisks indicate significance levels for the interaction (see Table 2 for details): *<.05; **<.005; ***<.001.
Figure 8
Figure 8
Top: Mean BOLD responses (in PSC units; see Fig. D1 in Fedorenko et al., 2010, for more details) in the language ROIs to word lists (red bars) and jabberwocky sentences (blue bars) relative to the fixation baseline. The ROIs were defined by intersecting the parcels, whose outlines are shown in grey (see also Fig. 5, bottom), with subject-specific thresholded (at p<.05, FDR-corrected) activation maps for the Sentences>Nonword-lists contrast, as described in Fedorenko et al. (2010). Each of these regions replicates the Sentences>Nonwords effect in independent data (left out runs), but shows no difference in response between word lists and jabberwocky. Bottom: The results of random-effects group analyses for the Word-lists>Jabberwocky contrast (red) and the Jabberwocky>Word-lists contrast (blue). The activation maps are thresholded at p<.05, FDR-corrected. (For the W>J contrast, no voxels emerged at this threshold. As noted in the text, for the J>W contrast, only the inferior posterior temporal / occipital regions – that fall outside of the classical language network – emerge at this threshold.)
Figure 9
Figure 9
The results of F-tests evaluating which regions can discriminate between word lists and jabberwocky sentences in the pattern of activity. Top: The results of ROI-based analyses. We show uncorrected values here (see Table 1 for details). If no value is shown for an ROI, this means that the region does not discriminate significantly above chance between the two conditions. Bottom: The results of the searchlight analyses at two different thresholds: .001, uncorrected, and .01, uncorrected.

Similar articles

See all similar articles

Cited by 36 articles

See all "Cited by" articles

Publication types

Feedback