Massaro et al. (1991) have shown that orthographic and semantic cues can jointly provide better memory retrieval than either cue alone (in a process called multiplicative cuing). Most commonly, this result is experienced by individuals who solve crossword puzzle on a regular basis. According to Nickerson (1977), crosswords can be viewed as cued retrieval tasks containing both orthographic cues (i.e., the letters already printed in the grid) and semantic cues (i.e., the Across and Down clues). Other crossword studies have further suggested that proficiency of puzzles depends more on the manipulation of letters and word fragments rather than a knowledge of a large number of words (Underwood, Diehim, and Batt, 1994). According to Hambrick, Salthouse, and Meinz (1999), "there is no indication . . . that either inductive or analytical reasoning processes contribute to the success in crossword puzzle solving." Instead, word retrieval remains central to the task, and the question remains: do certain arrangements of letters-like those frequently encountered in a crossword puzzle-constitute a distinct unit in the mental lexicon?

Since the early nineteenth century, many linguistic and cognitive neuroscientists have argued that some such units may include the phoneme, the morpheme, the diphthong, and the syllable. While all these classifications have been hard to define, the delineation of the syllabic function has been particularly so. According to the third edition of the American Heritage Dictionary of English Language, a syllable is "a unit of spoken language consisting of a single uninterrupted sound formed by a vowel, diphthong, or syllabic consonant alone, or by any of these sounds preceded, surrounded, or followed by one or more consonants."

Syllables are widely accepted as important for both auditory word perception (Kahn, 1976; Mehler, et al., 1981; Segui, 1984) and written word recognition (Goldblum and Frost, 1988). In this last study, participants were asked to generate words from a series of word fragments, accompanied by semantic clues. Ultimately, the participants retrieved words more accurately and quickly from word-initial syllabic fragments (e.g., M E R _ _ _ _ for MERCURY) than from non-initial pronounceable clusters (e.g., _ E R C _ _ _ ), non-initial unpronounceable clusters (e.g., _ _ R C U _ _ ), or randomly dispersed letters (e.g., M _ R _ U _ _ ). This suggests that organization of letter strings, alone, is not enough to facilitate word retrieval. Lexical units seem to be organized according to phonological principles-in this case, the syllable.

Only a few positron emission tomography (PET) studies and functional magnetic resonance imaging (fMRI) studies have examined lexical units. Passive listening or silent reading of words have been shown to activate Brodmann's areas 22/42 in the left temporal cortex (Peterson et al., 1990; Frith et al., 1991; Howard et al., 1992; Price et al., 1997). Discriminating between concrete and abstract words activates the posterior part of the left superior and middle temporal gyrus (Friederici et al., 2000) and discriminating between spoken pragmatic, semantic, and syntactic sentence errors activates distinct brain regions to varying degrees (Kuperberg et al., 2000). In another recent fMRI study, German participants were flashed twenty unpronounceable letter strings and twenty highly imaginable German nouns (Jessen et al., 1999). Classical language-related brain areas such as the left inferior frontal gyrus, the left superior temporal gyrus, and the left parietal-occipital regions were activated with the random letter strings, alone. The whole German words, however, activated additional areas: the left angular gyrus, the bilateral percuneus and the left posterior cingulate gyrus. In short, the study suggests that letter strings alone can activate language areas, but other regions may be reserved for higher semantic associations such as the word. Does the syllable, however, activate an amount of brain region somewhere between that of the letter strings and the whole word? Or does it activate the same minimal region as the letter strings or the same maximal area as the whole word? Perhaps the syllable can even activate a unique region.

The goal of this paper is to answer these hypothesis questions using a series of syllabic cues, typical to those encountered in a crossword puzzle. Semantic cues are also added at varying stages to test the neural basis of multiplicative cuing.

Methods

Participants.

Four healthy Duke undergraduates-three males and one female-were recruited through word of mouth. None of the participants, ages 18- to 22-years-old, had a history of neurological injury or disease such as trauma, seizures, strokes, or prior neurosurgery. Three of the four participants were right-handed, and all were monolingual English speakers. Each participant was compensated for his time. This study was approved by the Duke University Medical Center Institutional Review Board, and each participant provided informed consent.

Stimulus Display.

Each participant was tested for 40 words and 5 practice words, each presented across four consecutive stages (see Table 1). All words were eight letters in length, began with a three-letter syllable (consonant-vowel-consonant), and fell within a low frequency range (0 - 45) according to Kucera and Francis (1967). Whenever possible, the fourth letter of each word was also a consonant to ensure a clear division between the first and second syllable. All letters were white and appeared on a black background. The participants viewed the displays through an angled mirror mounted on eyeglasses.

Each word set was divided into four stages. In the first stage, participants were asked to determine whether a random 3-letter string of letters was in alphabetical order. The three-letter strings were placed either in the first three positions of the mock clue (e.g., t a s _ _ _ _ _ ) or in the third, fourth, and fifth positions (e.g., _ _ w e r _ _ _ ). Care was taken not to pick three letters that were actually part of the 40 test words, thereby minimizing the influence of priming. Underneath each clue fragment, the word "alphabetize" appeared to remind participants of the task. This stage served as a control to Stage 2.

In the second stage, participants were asked to think of a word from a three-letter string. When placed at the beginning of the clue, the three-letter string was a syllable

(e.g., c o n _ _ _ _ _ for concerto). When placed in the third, fourth, and fifth positions, the three-letter string was a nonpronounceable cluster (e.g., _ _ n c e _ _ _ for concerto). From the list of 40 test words, participants were presented with an equal number of alternating syllables and nonprounceable clusters (see Table 2). Underneath each clue fragment, the word "fill-in-the-blank" appeared to remind participants of the task.

In the third stage, participants were asked to think of a word from the same three-letter string encountered in stage 2; this time a semantic clue was printed underneath the word fragment. Letter position was conserved. So, if a participant was presented with c o n _ _ _ _ _] in Stage 2, they would be presented with c o n _ _ _ _ _ and the clue "masterpiece." If they were instead presented with _ _ n c e _ _ _ in Stage 2, they would be presented with _ _ n c e _ _ _ with the clue "masterpiece." As such, participants were presented with an equal number of semantic clues with syllabic prompts (n = 20) and semantic clues with cluster prompts (n = 20).

In the final stage, the words "say answer" were flashed on the screen, giving the participant an opportunity to communicate his answer for the word fragment.

Procedure.

Before the experiment began, each participant were given directions and asked to complete five practice sets, generated on the computer using Showtime software (The MathWorks Company).

All scanning was performed on a General Electric 1.5T scanner equipped with an Advanced Development Workstation for realtime echoplanar imaging. The experimenter identified the location of the anterior and posterior commisure, and 16 slices on this plane, 7.5mm thick, were chosen for the study. Sagittal T1-weighted localizer images were first collected for approximately 30 minutes. The functional images measured changes in BOLD contrast and were collected using a T2*-weighted gradient-echo, echoplanar imaging sequence [echo time (TE) = 40 msec; repetition time (TR) = 1.5 s; matrix = 64 x 64; field of view = 24 x 24 cm].

The 40 test words were divided into 4 runs of 10 word sets each, allowing participants to have short mental breaks in between. During the runs, the alphabetizing stage was first presented for 3.0 s, and participants were asked to indicate with a two-button system whether or not the letter-strings were in alphabetical order. The responses were recorded in a computerized output file. After this stimulus, 6.0 s of blank screen was presented to ensure recovery of the hemodynamic response (Huettel and McCarthy, 2000). Afterwards, the fill-in-the-blank task was presented for 10.5 s followed immediately with the fill-in-the-blank task with a semantic clue for another 10.5 s. Immediately afterwards, the screen "say answer" was presented for 3.0 s, and participants were given an opportunity to verbalize their response. The correct answer was then flashed for 1.5 s. Finally, 6.0 s of blank screen was flashed to ensure recovery of the hemodynamic response before the next word was tested (see Table 5 for schematic illustration). One word set then lasted 40.5 s. With a 12 s delay at the beginning of each run, each run lasted 417 s (6 min, 57 s).

Table 5.

In order to ensure that words were tested for both syllables and nonpronounceable clusters, two versions of the test were alternated between subjects. For example, participant 1 and 2 were tested for "concerto" in the syllable format (i.e., c o n _ _ _ _ _ ) while participants 3 and 4 were tested in the nonprounceable format (i.e., _ _ n c e _ _ _ ). For the next word, the formats would be switched-participants 1 and 2 would receive the nonpronounceable format and participants 3 and 4 would receive the syllable formats (see Table 2). This variation eliminates the small possibility that "easier" words were picked for the syllable format and "harder" words were picked for the nonpronounceable format. In short, if there is a difference between formats, the reason does not lie in the word selection.

Data Analysis.

The statistical analysis of the fMRI data was performed at the individual level (within-subject analysis). The data was analyzed by using a subtractive t-test to compare activation from one condition with the activation from another condition. First, the activation from the alphabetizing task was compared to activation invoked by the fill-in-the-blank task. Therefore, the t-test eliminated the similar areas of activation that these two tasks shared. Then, the activation from the fill-in-the-blank task was compared to the activation from the fill-in-the-blank task plus semantic clue. Finally, the activation from syllable processing was compared to differential activation from the nonpronounceable cluster tasks. The Matlab Overlay2 function (The MathWorks Company) was utilized to project the acquired activation as a Z-score map superimposed over anatomical brain images. The significant Z value was thresholded at a value of 3.6 which correlated to a p value of .001. Activated brain regions were identified using the Talairach and Tournoux brain atlas (1988). The Region of Interest (ROI) Matlab function was used to analyze activation common to all subjects. Due to time constraints, only one subject was analyzed with ROI. These calculations were performed using the bin analysis of epochs acquired during the scan.

The alphabetize task significantly activated (p < 0.001) a distinct region of the visual cortices in 3 out of 4 participants (Table 3; Figure 1a), and the hemodynamic response of Participant 2 (Figure 4a) is consistent with the 6 s expected delay. Here we see activation beginning 6 s after the onset of a stimulus and peaking around 7 s. The fill-in-the-blank task significantly activated (p < 0.001) distinct regions of the visual cortices in 4 out of 4 participants. These regions were located posteriorly and laterally to the visual regions uniquely activated by the alphabetize task (Figure 1b). The fill-in-the-blank task also significantly activated (p < 0.001) a region of Broca's area in 4 out of 4 participants (Figure 1c). The hemodynamic responses of Participant 2 to these fill-in-the-blank regions, however, (Figure 4b & 4c) are suspect and need to be further examined (see Discussion). Nevertheless, our results are largely consistent with Jessen et al. (1999): the fill-in-the-blank task required higher semantic associations compared to the random letter strings.

Table 3: Voxel coordinate comparisons from significant regions of activation in alphabetize vs. fill-in-blank tasks.

Figure 1a: Alphabetize task (red) vs. fill-in-the-blank task (blue) Alphabetize task activates visual cortices in 3 out of 4 participants.

Figure 1b: Alphabetize task (red) vs. fill-in-the-blank task (blue) Fill-in-the-blank task activates higher order visual cortex in 4 out of 4 participants.

Figure 1c: Alphabetize task (red) vs. fill-in-the-blank task (blue) Fill-in-the-blank task activates Broca's area in 4 out of 4 participants.

Figure 4a: Hemodynamic response of Participant 2 in visual cortex for alphabetize task

Figure 4b: Hemodynamic response of Participant 2 in visual cortex for fill-in blank task

Figure 4c: Hemodynamic response of Participant 2 in Broca's area for fill-in blank task.

The fill-in-the blank task with semantic clues activated distinct brain regions compared to the fill-in-the-blank task, alone.

The fill-in-the-blank task without a clue significantly activated (p < 0.001) a distinct region of Broca's area in 2 out of the 4 participants (Table 4; Figure 2a). However, the fill-in-the-blank task with clues significantly activated (p < 0.001) a distinct region of Broca's area in 4 out of 4 participants (Figure 2b). The hemodynamic response of Participant 2 (Figure 4d) to the fill-in-the-blank with clue task is suspect and needs to be further examined (see Discussion). Nonetheless, our results are consistent with Massaro et al. (1991) and Jessen et al. (1999): semantic clues will assist in word retrieval and activate higher semantic associations.

Syllables do not activate additional brain regions compared to nonprounceable clusters.

Since the fill-in-the-blank task resulted in different regions of activation for the clue and non-clue conditions (see above), we separated these conditions in our analyses of the syllables and nonprounceable clusters. However, contrary to our initial hypotheses, syllables did not activate distinct brain regions compared to nonpronounceable clusters in either the fill-in-the-blank task without clues (Figure 3a) or the fill-in-the-blank task with clues (Figure 3b). These results raise into question the importance of distinct neural substrates for seemingly different cognitive cues.

Figure 3a: Syllables (red) vs. clusters (blue) for fill-in-the-blank without clues. Syllables and clusters do not activate distinct regions in the fill-in-the-blank task without clues.

Figure 3b: Syllables (red) vs. clusters (blue) for fill-in-the-blank with clues. Syllables and clusters do not activate distinct regions in the fill-in-the-blank task with clues.

In light of this model, our alphabetize task distinctly activated regions of the primary visual cortex whereas the fill-in-the-blank task distinctly activated higher order visual cortices, the parietal-temporal-occipital association cortex (Figure 5a), and regions of Broca's area (Figure 5b). As stimuli, the two tasks were visually the same (three letters and five blanks with a printed instruction underneath). However, the activated brain regions were profoundly different. As such, the cognitive task determined the level of activation. When asked to alphabetize a set of letters, participants only seemed to activate the first part of the Wernicke-Geschwind language loop. Higher-order semantic regions do not seem necessary for the alphabetize task. On the contrary, when asked to generate a word, participants activated higher levels of the Wernicke-Geschwind loop. This result suggests that the mere thinking of a word can activate regions of the brain similar to that of reading a word (as proposed by Wernicke and Geschwind). When just manipulating letters (as was the case with the alphabetize task), only lower regions of the Wernicke-Geshwind loop need to be employed.

Figure 5a: Fill-in-the-blank task activates a different region than the alphabetize task in the Wernicke-Geschwind loop of language processing.

Figure 5b: The fill-in-the-blank task activates Broca's area in the Wernicke-Geschwind pathway whereas the alphabetize task does not.

Within the fill-in-the-blank task, another result was evident from our study. The addition of a semantic clue activated a distinct region of Broca's area compared to the fill-in-the-blank tasks without clues. Visually, both of the stimuli appeared the same (one had "fill-in-the-blank" written underneath, the other had the clue printed below). However, the addition of a semantic cue prompted the use of a different region of Broca's area. Therefore, we present for the first time anatomical evidence to support Massaro et al.'s multiplicative theory (1991). In this theory, semantic and orthographic clues are jointly predicted to provide better memory retrieval than either cue alone. That the addition of a semantic clue activates a distinct region of Broca's area suggests that distinct neural substrates are employed. Additionally, the fact that this distinct region also lies within Broca's area is testament to the capacity of Broca's area to integrate multiple linguistic inputs into cohesive units of comprehension.

Finally, our results showed no differences in word retrieval between syllabic and nonsyllabic cues. Based on our particular task, then, we can offer no evidence to suggest that different neural substrates are employed for syllabic cues and nonsyllabic cues. However, our results can neither negate the hypothesis that the syllable is a unit of the mental lexicon. Previous cognitive studies have already shown that syllabic cues prompt better word retrieval than nonsyllabic prompts (e.g., Goldblum and Frost, 1988). Our failure to find distinct neural substrates could be (1) because the word tasks are not difficult enough to see a difference or (2) syllables and nonsyllables do not employ the same language areas but, instead, activate regions in a different manner. For example, syllables and nonsyllables could activate all the same regions of Broca's area but in a slightly different order. Such minor differences are not likely to be captured through a functional magnetic imaging study. But alternatively, it could be true that no anatomical difference exists between syllabic and nonsyllabic cues. If such were to be the case, then previous cognitive studies would need to be re-examined in light of potential priming effects, and the syllable would need to be re-examined as a distinct unit of the mental lexicon.

With this said, our results also provide some insight into the famous neurological patient H.M. who underwent bilateral medial temporal lobe resection for the relief of intractable epilepsy in 1953 (Milner, 1972; Scoville and Milner, 1957). The resection included the medial temporal polar cortex, most of the amygdaloid complex, most or all of the entorhinal cortex, and approximately half of the rostrocaudal extent of the intraventricular portion of the hippocampal formation (Corkin et al., 1997). As a subsequent result of this surgery, H.M. displayed massive anterograde amnesia and was among the first to reveal the role of temporal-love structures on human memory (Corkin, 1984; Milner, Corkin, and Teuber, 1968). Perhaps greatest among these findings was the dichotomy of H.M.'s semantic memories. Since the surgery, H.M. has displayed normal memory for semantic knowledge acquired before 1953, but he has been profoundly amnesic for semantic knowledge acquired after 1953 (Gabriele, Cohen, and Corkin, 1988).

Interestingly, however, H.M. retained his lifelong love of crossword puzzles. According to his own verbal reports, he began solving crosswords that were printed in his local newspaper around the age of 15 (Skotko et al., in review). Retention of such an interest is rather surprising, though, since crossword puzzles tend to rely heavily on current events and popular fads. Nevertheless, H.M. continues to work on two or more puzzles each day despite the fact that he has a severe amnesia. Even those who are close to him indicate that he frequently works on his crossword puzzles. He even engages in challenging puzzles featured in books published by the N.Y.Times and Merriam Webster .

Of course, H.M. has demonstrated a learned retention of perceptual-motor skills. In Milner's (1962) experiment, he was able to learn a mirror-tracing task across three days of testing. Although there was no indication in his verbal reports that H.M. remembered each day's experience, this study suggests that with practice, he was able to derive the solution to this puzzle.

Crossword puzzles, however, are more complex than simple collections of perceptual-motor skills. General knowledge, word retrieval skills, and crossword puzzle experience, for example, have been shown to be major predictors of puzzle proficiency (Hambrick, Salthouse, and Meinz, 1999). Together, these variables accounted for approximately 85 per cent of the variance in puzzle success. Witte and Freund (1995) also showed that performance on anagrams-another independent test of word retrieval skills-was positively correlated with previous puzzle experience.

Regardless of the many predictors, however, the use of reasoning skills remained surprisingly absent from all sets of variables. According to Hambrick, Salthouse, and Meinz (1999), "there is no indication . . . that either inductive or analytical reasoning processes contribute to success in crossword puzzle solving." While such a finding is certainly intriguing for normal individuals, it is particularly so for H.M. who has largely retained his procedural memory (Milner, 1962). If this component of memory does not play a large role in puzzle proficiency, however, H.M. must rely on some form of semantic memory to solve the clues. Yet, his postoperative semantic memory is severely impaired. Why, then, does H.M. have such a fascination with these crossword puzzles?

The results from our study suggest that higher-order visual cortices, parietal-temporal occipital association cortices, and regions of Broca's area are all employed in our crossword-like tasks. These areas are preserved in H.M. and could perhaps explain his retained ability and interest in solving crossword puzzles. If reasoning and a large vocabulary do not play an integral role in the solution of these puzzles (as the current research indicates), then it should be expected for HM to be a competent puzzle solver. Due to his lesions, only those clues requiring postoperative memories should cause difficulties.

Finally, some words about the study's limitations are in order. Participant 1 and 2 appeared to move while in the scanner, resulting in a lot of "popcorn effects" (Figures 1a and 1b). Ideally, these participants would be discarded from a study; but based on our limited time and availability of participants, they were included in this analysis. Additionally, the hemodynamic responses need to be further explored. Due to time constraints, we only were able to calculate the response for Participant 2; further analysis should be done on the remaining Participants. For participant 2, the response to the alphabetize task appears normal (Figure 4a), while the response for the fill-in-the-blank has a disturbing peak between 11s and 13s (Figure 4b). This peak could be due to the subsequent solving of the previous alphabetizing task-i.e., perhaps, the participant was still thinking of the previous task and actually solved it during this time. The rise around 15 s appears to be the normal response for the fill-in-the-blank task. The hemodynamic response in Broca's area, however, (Figure 4c) could be complicated by different solution times. The trimodal response could suggest that the participant was likely to solve the clues in three time frames: either right away, approximately 2 s thereafter, or after about 10 s of thinking. Once the participant solved the clue, we would expect the hemodynamic response to drop as he or she would likely cease to think about solving the clue. That the response is trimodal could point to the varying degrees of difficulty in our clues. Similarly, the hemodynamic response for the fill-in-the-blank task with clues (Figure 4d) demands a similar explanation. Further tests should be done to explore these hemodynamic responses.

We also realize that our sample size (n = 4) is small, although we believe that this small number of subjects was adequate for a preliminary study. We are aware that the stimuli that were utilized may have been interpreted differently by the various subjects and had the potential to be more complex than intended. It is therefore possible that associations were made between the various stimuli. This leads to the possibility that certain responses or activated areas of the brain may have been due to something other than the single stimuli utilized at the time of evaluation. We are aware that the hemodynamic responses noted in a limited number of subjects may not be consistent as larger numbers of subjects are examined. However, the responses that we were able to demonstrate will at least serve as a basis for future research in this area.

In all cases, the same brain regions were activated in at least three out of the four participants. That alone is preliminary evidence to suggest that the regions are indeed distinct and real. Furthermore, the anatomical evidence of Massaro's multiplicative theory of orthographic and semantic cues highlights the minor distinctions that need to be made between seemingly similar language tasks. Future functional magnetic imaging studies should be cognizant of the significant differences in activation that could result from minor adjustments in the cognitive task.