How readers encode and process morphologically complex words has become a central issue in cognitive psychology, especially in research on visual-word recognition. Recognizing an isolated polymorphemic word involves a number of underlying processes: encoding of letter position/identity, morphological segmentation, lexical retrieval of each morpheme, and whole-word semantic information retrieval. The precise manner and order in which these processes occur (or co-occur) is a matter of contention that is subject to empirical research (see Amenta & Crepaldi, 2012, for a comprehensive review).

Most researchers in visual-word recognition agree that the recognition of derived words (e.g., violinist) is guided by early morpheme detection processes occurring before whole-word processing has finished (violin + ist; e.g., Pastizzo & Feldman, 2002, 2004; Rastle, Davis, & New, 2004; Taft & Forster, 1975, 1976; see also Diependaele, Duñabeitia, Morris, & Keuleers, 2011, for evidence from nonnative speakers). Empirical evidence in favor of this view has generally been obtained from masked-priming experiments, in which the processing of a briefly presented prime is posited to reveal early nonstrategic processes (see Diependaele, Morris, Serota, Bertrand, & Grainger, 2013, for review). Consequently, most studies on derivational morphology have concluded that morphological decomposition occurs at a very early stage of visual-word recognition (see Lavric, Elchlepp, & Rastle, 2012).

Thus, one relevant question is whether (some of) the processes of morphological decomposition co-occur with letter position coding at very early stages of visual-word recognition. Following this line of reasoning, Christianson, Johnson, and Rayner (2005), conducted a masked-priming experiment (Exp. 3) examining whether letter transpositions across affix boundaries (e.g., boas et r–BOASTER) differed from transpositions that crossed a pseudo-affix boundary in a naming task (e.g., blus et r–BLUSTER; note that blust is not an English stem, as -er is not a real suffix in bluster). Transposed-letter (TL) effects are prototypical orthographic effects associated with letter-position coding mechanisms (e.g., Perea & Lupker, 2003; see Gomez, Ratcliff, & Perea, 2008; Grainger & Whitney, 2004, for theoretical implications). Christianson et al. found that “TL primes did not significantly differ from the orthographic controls” (p. 1334; in their case, one-letter substitution primes—e.g., boas l er–BOASTER) in the set of truly suffixed words. That is, naming latencies for suffixed words preceded by a TL nonword prime that crossed the affix boundary were similar to latencies for suffixed words preceded by an orthographic control (both ts < 1; see also Luke & Christianson, 2013, for related findings). Nonetheless, the statistical significance of the results was not fully unambiguous, and these data should be taken as suggestive rather than conclusive.

Stemming from these results, Duñabeitia, Perea, and Carreiras (2007) extended the finding that masked TL priming was affected by early morphological effects. In a series of lexical decision experiments with Basque and Spanish speakers, Duñabeitia et al. (2007) found significant masked TL priming effect for polymorphemic words when the letter transposition occurred within a morpheme (vio il nist–VIOLINIST), whereas they failed to find a parallel effect for affixed words when the transposed letters crossed the affix boundaries (violi in st–VIOLINIST vs. violi er st–VIOLINIST). To explain these findings, the authors suggested the existence of a morpheme detection mechanism operating “early in the process of visual word recognition, co-occurring with mechanisms responsible for assigning letter position” (p. 701; see also Duñabeitia, Perea, & Carreiras, 2008), akin to the prelexical affix-identification mechanism proposed by Taft and Forster (1975, 1976).

However, a recent experiment by Sánchez-Gutiérrez and Rastle (2013) directly disputed the Duñabeitia et al. (2007) findings. The authors reviewed other studies testing English participants in which the magnitudes of TL priming effects were similar in size, regardless of whether or not the transpositions crossed the morphemic boundary (e.g., Beyersmann, Coltheart, & Castles, 2012; Beyersmann, McCormick, & Rastle, 2013; Masserang & Pollatsek, 2012; Rueckl & Rimzhim, 2011). They then questioned whether the discrepancy between the data reported by Duñabeitia et al. (2007) and these studies could result from an inherent difference between the languages at test. To test their hypothesis, Sánchez-Gutiérrez and Rastle selected a set of 88 derived words that were cognates between English and Spanish (viz., words with highly similar orthographic, phonological, and semantic representations, such as invisible), and tested English and Spanish participants in a masked-priming TL experiment with a lexical decision task in which the manipulations took place either at the morphemic boundary (i.e., between morphemes) or within morphemes. Sánchez-Gutiérrez and Rastle found masked TL priming effects of similar magnitude for between- and within-morpheme transpositions, as deduced from the lack of a significant interaction. They concluded that “the findings of Duñabeitia et al. reflected idiosyncratic properties of the stimuli or the participants, or a Type I error” (p. 992). Certainly, one of the risks in psycholinguistic studies testing a relatively small set of items is the appearance of a false positive (Type I error). Of course, an identical argument applies to the possibility of reporting a false negative (Type II error), in particular when failing to detect an interaction when the expected effect size is very small.

What is the reason for these incongruent results? Leaving aside issues related to statistical power or to the potential particularities of the stimuli used in the previous experiments, the diverging results for TL manipulations across morphemic boundaries could potentially stem from individual differences in reading skills on morphological decomposition and morpho-orthographic interactions. In a compelling study on morphological priming on words with transparent and opaque relationships (i.e., walker–WALK vs. corner–CORN), Andrews and Lo (2013) demonstrated that the participants’ reading styles (orthographic-based vs. semantic-based) modulated the magnitude of these effects, especially in the case of opaque relationships—note that these results have been considered as a marker of morpho-orthographic processing. Although good spellers with average vocabulary (i.e., readers with an orthographic profile) showed significant priming effects for opaque pairs, high-vocabulary participants with average spelling skills (i.e., readers with a semantic profile) showed minimal priming for opaque relationships. Andrews and Lo concluded that the “individual differences approach provides critical, novel evidence that explains the contradictions observed in published average data and provides new evidence that contributes to distinguishing the relative validity of theories of morphological priming” (p. 289).

To examine in detail whether morphological decomposition in suffixed words interacts with letter-position encoding, we conducted a large-scale masked-priming lexical decision experiment including transposed-letter manipulations between morphemes (across morphemic units) and within morphemes. Large-scale experiments partially solve the important issue of statistical power (Keuleers, Diependaele, & Brysbaert, 2010; see also Button et al., 2013). To this end, we selected a set of 420 suffixed Spanish words and tested 80 participants. Given the relatively large size of the sample, and considering that individual differences in reading skills modulate the extent to which a morpho-orthographic style of processing is followed (see Andrews & Lo, 2013), we divided the group of participants according to their reading speeds (using a median split), and examined whether individual differences could be responsible for the seemingly discrepant results reported in recent years. Previous psycholinguistic research has shown a close interrelation between variations in orthographic processing skills and individual differences in reading (e.g., Perfetti & Hart, 2001; Stanovich & West, 1989; see Perfetti, 2007, for review). As was recently reported by Hargreaves, Pexman, Zdrazilova, and Sargious (2012), orthographic skills are inversely correlated with reading speed in the lexical decision task (viz., shorter reaction times [RTs] for participants with high orthographic skills; see also Chateau & Jared, 2000, and Unsworth & Pexman, 2003, for a similar argument regarding the influence of increased sublexical skills on reading latencies), and they are also inversely correlated with semantic effects (e.g., participants with enhanced orthographic abilities show reduced effects of concreteness). Hence, following the hypothesis developed by Andrews and Lo regarding the close link between orthographic skills and morpho-orthographic interactions, on the one hand, and considering the evidence suggesting that readers with higher orthographic skills show faster RTs in the lexical decision task than do readers with lower orthographic skills, on the other hand, we expected to find that morpho-orthographic interactions were modulated by reading speed.

Method

Participants

A group of 80 undergraduate and graduate students (49 females, 31 males) took part in this experiment. All of them were native Spanish speakers, had normal or corrected-to-normal vision, reported no history of neurological disorder, and signed informed consent forms before the experiment. Their mean age was 25.47 years (range: 18–40, SD = 5.94).

Materials

A set of 420 Spanish suffixed words was selected from B-Pal (Davis & Perea, 2005). The characteristics of these words are reported in Table 1, and the full list of materials is presented in Appendix A. The set of words included 25 different suffixes, thus comprising a representative subset of Spanish affixed words. These words acted as targets (e.g., VIOLINISTA, violin + ista; translated as violinist) and could be preceded by a prime that was (1) a nonword created by transposing two adjacent letters that crossed the morphemic boundary (TL-between condition; e.g., violi in sta); (2) a nonword created by replacing those two transposed letters with others (replaced-letter [RL]-between condition; e.g., violi er sta); (3) a nonword created by transposing two adjacent letters that did not cross the morphemic boundary (TL-within condition; e.g., vio il nista); or (4) a nonword created by replacing those two transposed letters with others (RL-within condition; e.g., vio at nista). As in previous research (e.g., Sánchez-Gutiérrez & Rastle, 2013), we restricted the criteria for replacements, minimizing the variability in height (i.e., trying to preserve the amount of vertical space occupied by each transposed letter), and replacing letters as a function of their consonant–vowel status. None of the letter transpositions or replacements involved two vowels (see Perea & Acha, 2009, for a demonstration of weaker TL effects for vowel combinations than for manipulations involving consonants) or the initial or final letters of the strings, and all of the bigrams manipulated resulted in existing letter combinations in Spanish (see Frankish & Turner, 2007). Special attention was paid to the frequencies of the manipulated bigrams, which were matched across TL and RL conditions in a pairwise manner: TL-between = 2,210.63, RL-between = 2,250.59, paired samples t(419) = –1.11, p = .27; TL-within = 1,639.22, RL-within = 1,625.74, paired samples t(419) = 0.32, p = .75 (see Frankish & Barnes, 2008, and Perea & Carreiras, 2008, for evidence regarding the importance of matching the different conditions in their bigram frequencies). Additionally, a set of 420 Spanish nonword targets was created (e.g., DULEFO; see Appendix A). We used Wuggy (Keuleers & Brysbaert, 2010) to create these nonwords from the real words; this program provides a valid matching of the items in subsyllabic structure and transition frequencies. These nonwords were matched to the words in length in a pairwise manner, they respected Spanish orthotactics, and they had bigram frequencies and orthographic neighborhoods similar to those of the word set (for the words: mean length = 8.97, mean type bigram frequency = 369.24, mean token bigram frequency = 77.11, mean N = 0.37; for the nonwords: mean length = 8.98, mean type bigram frequency = 341.69, mean token bigram frequency = 70.09, mean N = 0.04). Considering the specific characteristics of the matching algorithm used by Wuggy, some nonwords ended in a sequence of letters that matched a real Spanish suffix, and others did not. Half of the nonword targets were preceded by a nonword prime that included an internal adjacent transposition, and half were preceded by a nonword prime that included a double-letter replacement. The manipulations carried out in the nonwords were distributed among different string (internal) locations, in order to emulate the manipulations in the word set. All of the word targets appeared in each of the four lists that were created, but each time in a different priming condition. Twenty participants completed each of the four lists, following a counterbalanced design.

Table 1 Characteristics (means, standard deviations [SDs], minimum values, and maximum values) of the different indices of the words used in the experiment
Full size table

Procedure

The experiment was conducted in acoustically shielded individual test cabins, using PCs (Dell Optiplex 760) with CRT monitors working at 1,024 × 768 and 90 Hz. Stimulus presentation and data collection were controlled by DMDX software (Forster & Forster, 2003). Each trial consisted of the centered presentation of a mask (# symbols) for 500 ms, followed by brief presentation of the prime in lowercase Courier New font for 55 ms (five cycles of 11.11 ms each). After this, the target appeared in uppercase letters and stayed on the screen for 2,500 ms or until the participant responded. The length of the mask varied from trial to trial, depending on the number of characters in the primes/targets. Participants had to press one of two labeled buttons on an Empirisoft DirectIN High Speed Button-Box, to indicate whether the displayed string was or was not an existing word in Spanish. They were instructed to do so as quickly and accurately as possible, and they were trained with a short practice consisting of ten words and ten nonwords. All of the experimental items were randomly presented. The whole experimental session lasted approximately 20 min.

Results

Trials associated with erroneous responses (3.94 %) and response latencies faster than 250 ms or slower than 1,500 ms (3.37 %) were excluded from the RT analysis. Then, a median-split procedure was followed in order to identify faster and slower participants (see Häikiö, Bertram, Hyönä, & Niemi, 2009, for a similar procedure). The mean RT for each participant across word and nonword trials was calculated, and individual participants were then split into two groups, depending on whether their mean RT fell above (i.e., slower group) or below (i.e., faster group) the median RT value of the averaged latencies for the whole set of items (median = 741 ms, mean = 761 ms, SD = 144 ms). The slower group (N = 40) had a mean overall RT of 881 ms (SD = 94 ms), whereas the faster group (N = 40) had a mean RT of 641 ms (SD = 62 ms). Next, participant- and item-based analyses of variance (ANOVAs) were run on the word data following a 2 × 2 × 2 × 4 design, including the factors Place of Manipulation (within | between), Type of Manipulation (TL | RL), Group (slower | faster), and List (1 | 2 | 3 | 4); List was included as a dummy variable in the design (Pollatsek & Well, 1995). The mean RTs and percentages of errors for each group and condition are presented in Table 2.

Table 2 Means and standard deviations of the lexical decision times (in milliseconds) and percentages of errors (within parentheses) for the word targets across the different conditions and groups
Full size table

The ANOVAs on the RTs for the word trials showed a main effect of place of manipulation [F 1(1, 72) = 43.82, p < .001; F 2(1, 416) = 58.52, p < .001] and a main effect of type of manipulation [F 1(1, 72) = 32.98, p < .001; F 2(1, 416) = 36.05, p < .001]. The effect of group was also significant [F 1(1, 72) = 170.22, p < .001; F 2(1, 416) = 5,856.25, p < .001]. Critically, the three-way interaction was significant [F 1(1, 72) = 9.46, p = .003; F 2(1, 416) = 6.80, p = .009]. Separate analyses were conducted for the faster and slower groups, to understand the origin of this interaction. For the faster participants, we observed a significant two-way interaction between place of manipulation and type of manipulation [F 1(1, 36) = 15.31, p < .001; F 2(1, 416) = 9.77, p = .002], reflecting a sizeable 18-ms TL effect for within-morpheme manipulations (TL = 597 ms, RL = 615 ms) [F 1(1, 36) = 49.84, p < .001; F 2(1, 416) = 34.18, p < .001], but not for between-morpheme manipulations (TL = 589 ms, RL = 593 ms) [F 1(1, 36) = 2.36, p = .13; F 2(1, 416) = 2.61, p = .11]. In contrast, the critical interaction did not approach significance for the slower participants [F 1(1, 36) = 1.37, p = .25; F 2(1, 416) = 1.09, p = .30]; note that both the place of manipulation [F 1(1, 36) = 15.12, p < .001; F 2(1, 416) = 20.40, p < .001] and type of manipulation [F 1(1, 36) = 12.81, p < .001; F 2(1, 416) = 16.19, p < .001] main effects were significant—the overall TL priming effect was 12 ms.Footnote 1

The analysis on the error data showed a main effect of place of manipulation [F 1(1, 72) = 22.01, p < .001; F 2(1, 416) = 18.00, p < .001] that interacted with group in the analysis by participants [F 1(1, 72) = 4.36, p = .04; F 2(1, 416) = 2.94, p = .09]. This revealed that within-morpheme manipulations led to higher error rates than did manipulations between morphemes, and that this difference was larger for the faster than for the slower group (a 1.30 % and a 0.45 % difference, respectively). No other effects/interactions were significant.

In sum, the present data reveal that for faster readers, TL priming effects were greater for within-morpheme than for between-morpheme transpositions, whereas no such difference occurred for the slower readers.Footnote 2 To obtain further evidence of the relationship between the overall RTs of the whole group of participants and the magnitudes of the TL effects within and between morphemes, two additional analyses were conducted. First, the magnitude of the TL effect for the within-morpheme manipulation was calculated for each participant (N = 80). The individual within-morpheme TL effects did not significantly correlate with the participants’ mean RTs [r = –.14, t(78) = 1.22, p = .22], showing that this effect was relatively independent from the processing speed. Importantly, the magnitude of the TL effect in the between-morpheme condition was positively correlated with the participants’ overall RTs [r = .26, t(78) = 2.40, p = .02], showing that the magnitude of the critical between-morpheme TL effect increased as a function of the general processing speed in the task (see Fig. 1).

Fig. 1
figure 1

Scatterplots of the mean transposed-letter (TL) effects (between morphemes [panel a] and within morphemes [panel b], respectively) for the whole set of participants as a function of the participants’ mean reaction times (RTs) in the lexical decision task

Full size image

Second, to further examine the manner in which the magnitude of TL priming effects is modulated by the participant’s rapidity in the task, we reexamined the data using a centile-based analysis of the RT distribution (see Andrews & Lo, 2013; Balota, Yap, Cortese, & Watson, 2008; and Yap, Tse, & Balota, 2009, for similar approaches; see also Ratcliff, 1979, for a description of the advantages of examining RT distributions over mean RTs, and Johnson, Staub, & Fleri, 2012, for further evidence regarding the relationship between reading speed and TL effects). We computed the RT distributions in each critical condition and obtained nine representative centiles (from the 10th to the 90th centiles, in steps of 10; see Fig. 2). Then we calculated the difference of the magnitudes of the TL effects for the between-morpheme and within-morpheme conditions for each participant. As can be seen in Fig. 2, the TL effects for within-morpheme manipulations were relatively unaffected by the speed of response, coinciding with the results from the correlational analyses. In contrast, the TL effects for between-morpheme manipulations increased as a function of the speed of response. ANOVAs performed on the whole set of data showed significant interactions between place of manipulation and type of manipulation in the 10th, 20th, 30th and 40th centiles (all Fs > 4, all ps < .05), whereas the interaction was negligible in the upper centiles (all Fs < 1, all ps > .30). Besides, as can also be observed in Fig. 2, the within- and between-morpheme TL priming effects were clearly different from each other in the group of faster readers, and this difference held constant across the entire RT distribution.

Fig. 2
figure 2

(a) Averaged reaction time (RT) distributions for the whole sample of participants (N = 80, central lines), and for the slow and fast groups (n = 40 each; top and bottom lines, respectively) in the transposed-letter (TL) and replaced-letter (RL) conditions (dotted and solid lines, respectively) for the between- and within-morpheme manipulations. (b) Scattering of the difference between the TL effects for the within- and between-morpheme manipulations according to the 10th- to 90th-centile distribution for the whole sample and for the slow and fast groups

Full size image

Discussion

The present large-scale masked-priming experiment helps reconcile earlier research on the interaction between early orthographic coding and morphological decomposition processes, by providing a tentative explanation for the divergent results reported in recent years. First, as Sánchez-Gutiérrez and Rastle (2013) observed, it is possible to obtain (fairly small) masked-priming effects when the transposed-letter manipulation crosses the affix boundary (around 10 ms, when averaging faster and slower readers), thus suggesting that the null effect reported in our earlier study (Duñabeitia et al., 2007) may have been the consequence of a Type II error (viz., a false negative). Second, and more importantly, when the individual differences of the participants in their reading speeds were taken into account, clear-cut differences emerged in the magnitudes of the TL priming effects for between- and within-morpheme manipulations. Specifically, even though the slower participants showed significant TL priming effects, irrespectively of the place of manipulation (hence replicating the results of Sánchez-Gutiérrez & Rastle, 2013, among others), the group of faster participants only showed TL priming effects in within-morpheme conditions, but not between morphemes (replicating Duñabeitia et al., 2007, among others). Hence, the present data demonstrate that the interaction reported by Duñabeitia et al. (2007) was not the result of a Type I error (a false positive), and that the null interaction effect reported by Sánchez-Gutiérrez and Rastle for the Spanish set is not a fully generalizable finding, given that it is dependent on the participants’ reading performance. These observations were further reinforced by correlational analyses and RT distributional analyses.

The present results add to a growing body of literature demonstrating the impact of individual differences in reading skills on the different stages of word processing (see Chace, Rayner, & Well, 2005; Häikiö et al., 2009; Janack, Pastizzo, & Feldman, 2004; Rayner, Slattery, & Bélanger, 2010; Yap, Tse, & Balota, 2009; Ziegler, Jacobs, & Klüppel, 2001, to cite just a few examples). Of particular interest is the recent study by Andrews and Lo (2013), in which the participants’ reading “profiles” (orthographic vs. semantic) modulated the magnitude of masked morphological priming effects—in particular, for morphologically opaque relationships (e.g., corner–CORN). The present data set and the different analyses reported validate Andrews and Lo’s argument in favor of studies testing individual differences in morphological priming, and they offer a tentative explanation for the seemingly inconsistent results reported in the literature on morpho-orthographic interactions. We found that the faster readers (who are, potentially, the best readers) are highly sensitive to morpho-orthographic interactions: When the transposed letters involved the final letter of the stem and the initial letter of the suffix in a polymorphemic word (e.g., “ni” in violi ni sta; i.e., a manipulation between morphemes), the magnitude of the TL effect was half the size of that of similar manipulations that exclusively involved two letters of the stem (e.g., “li” in vio li nista; i.e., within-morpheme manipulations). On the contrary, the slower readers failed to display a significant morpho-orthographic interaction.

The present data replicate and extend preceding studies that have suggested that minimal orthographic positional changes lead to manifestly different effects, depending on whether or not the morphological properties of the words are kept intact (Christianson et al., 2005, Exp. 3; Duñabeitia et al., 2007; see also Luke & Christianson, 2013, for additional evidence), and support a claim for the existence, at least for the most skilled readers, of an early stage of morphological decomposition that is sensitive to morpho-orthographic interactions (see also Taft & Nillsen, 2013). More importantly, the present study offers an (admittedly speculative, but suggestive) explanation for the different results that have been provided in recent studies on the same issue (see Beyersmann et al., 2013; Rueckl & Rimzhim, 2011; Sánchez-Gutiérrez & Rastle, 2013).

We suggest that TL effects across morphemes do not respond to a dichotomous all-or-nothing reality. Instead, TL effects across morphemic boundaries can be captured better by a continuum that may be linked to individual differences in reading performance and/or to participants’ reading styles/profiles. Considering the tight link between reading efficiency and orthographic knowledge (see, e.g., Perfetti, 2007), and taking into account recent evidence demonstrating that increased orthographic skills lead to shorter RTs in the lexical decision task (e.g., Hargreaves et al., 2012), we initially hypothesized that the faster participants would be the ones with an increased sensitivity to morpho-orthographic factors, and consequently the ones who would show minimal TL effects between morphemes. The results of the present experiment fully satisfied this hypothesis, suggesting that faster readers follow a reading style based on fast-acting automatic morphological decomposition processes (akin to the morpho-orthographic account proposed by, among others, Diependaele, Sandra, & Grainger, 2005). In contrast, the slower readers are hypothesized to focus less on orthographic information, probably basing their reading profile on semantic-based pieces of information (note, for instance, that the concreteness effect is inversely correlated with the orthographic skills of the readers, as was shown by Hargreaves et al., 2012). In this line, we found that slower readers showed no morpho-orthographic interactions and robust between-morpheme TL effects. In spite of the strength of the analyses reported here, we acknowledge that further studies should be aimed at reexamining the relationship between reading speed, orthographic skill, and morpho-orthographic interactions, probably using fully independent indices for each construct.

At a methodological level, we believe that one of the most efficient avenues to shed more light on the existence of morpho-orthographic interactions will be the implementation of masked-priming experiments testing: (1) a sufficiently large and heterogeneous sample of items that are representative of each language, and (2) a large sample of participants with varying degrees of reading skills. This joint approach may provide enough power to detect small differences that are inherent in subtle manipulations, as we have demonstrated in the present experiment. Following this strategy, the present data reinforce the view that although large effects in masked priming are reliable and consistent across labs, “anything finer-grained is extremely difficult to detect” (Gomez et al., 2008, p. 598), at least in typical experiments testing subtle effects with a relatively small set of words per condition. Furthermore, a large-scale approach would allow researchers to explore whether the individual characteristics of the words modulate the pattern of priming effects. In this light, we tested whether (and how) the specific properties of the words used in the experiment had an influence on the magnitude of the critical TL effect between morphemes in a set of regression analyses (see Appendix B). The results of these analyses did not highlight any relevant property of the words as being a modulating factor for slower readers. Similarly, the regression models for faster readers showed relatively poor fits, thus leading to the conclusion that the reason for the differences in the magnitudes of the TL effect is not “idiosyncratic properties of the stimuli,” as was suggested by Sánchez-Gutiérrez and Rastle (2013).

To sum up, the present data suggest that the process of letter position coding is not completely blind to the context in which the manipulations take place (e.g., between morphological units) or to the individual differences of the participants in their reading skills. Taking into account the data provided by Andrews and Lo (2013) and the present results, further research should try to clarify the influence of orthographic skill in the morphological decomposition of polymorphemic words.