Detecting spelling errors in compound and pseudocompound words

Acknowledgement: This research was supported by NSERC Discovery Grants (250028 and 05,100) to Christina L. Gagné and Thomas L. Spalding.

Intuitively, it seems that morphemes are recovered when they are productively used within a compound morphological structure (e.g., snow and ball function as morphemes within snowball) but not when they are only incidental and not involved in the morphological structure of a word (e.g., hip and pie do not function as morphemes in hippie). However, is this intuition correct? The aim of the current set of experiments is to investigate whether words are automatically segmented into morpho-orthographic units and, if so, whether the morphemic structure of the word influences the impact of the retrieved morphemes. In particular, we will examine whether the availability of morphemes differentially influences the processing of compounds and pseudocompounds (i.e., words that do not have a compound structure but can be split into two free morphemes). To do so, we evaluate the impact of embedded morphemes on the recognition of words for which those morphemes are used productively or not (i.e., have a false morphemic structure). We evaluate impact by measuring the ease with which people can decide whether the words contain spelling errors. In this article, we begin by discussing the theoretical issues concerning morphological decomposition and automatic segmentation before introducing a spelling error detection task as a way of evaluating the impact of morpheme availability on word recognition.

Although there has been substantial evidence to suggest that morphological representations become available during the processing of morphologically complex words (e.g., Andrews & Davis, 1999; Beyersmann, Castles, & Coltheart, 2011; Gagné & Spalding, 2004, 2009; Ji, Gagné, & Spalding, 2011; Libben, 1994; Sandra, 1994; Taft & Forster, 1975), there is still debate about when decomposition happens and about the impact that the constituent representations have on word access. Some researchers argue that access to morphological representations happens after the meaning of the whole word is accessed (e.g., Giraudo & Grainger, 2001; Giraudo & Voga, 2016; Levelt, Roelofs, & Meyer, 1999). However, others have suggested that decomposition happens earlier in processing, before the meaning of the whole word has been obtained, based on orthographic or morpho-orthographic representations (e.g., Crepaldi, Rastle, Coltheart, & Nickels, 2010; Taft & Forster, 1975). Findings supporting this approach come from masked priming lexical decision experiments (e.g., Beyersmann et al., 2016; see also Longtin, Segui, & Hallé, 2003, for similar findings in French; Diependaele, Sandra, & Grainger, 2005, for Dutch; Kazanina, Dukov-Zheleva, Geber, Kharlamov, & Tonciulescu, 2008, in Russian) that demonstrated that stem morphemes become available when a word contains a suffix, regardless of whether the word has a stem + affix morphological structure (e.g., hunter) or not (e.g., corner), but not when the word does not contain an affix (e.g., cashew). Primes with a true morphological structure (e.g., hunter-HUNT) or with a pseudomorphological structure (e.g., corner-CORN) equally aided the processing of the target, whereas nonsuffixed control words (e.g., cashew-CASH) do not aid processing.

Because the bulk of research on morphological decomposition has focused on affixed and pseudoaffixed words, it is still an open question as to whether words that contain embedded free morphemes but no affixes (namely, compound and pseudocompound words) also show similar effects. Therefore, these types of words will be the focus of the current investigation. Although compounds (e.g., necklace) appear to be the equivalent of affixed bimorphemic words such as hunter and pseudocompounds (e.g., carpet) appear to be the equivalent of pseudoaffixed mono-morphemic words such as corner, it does not directly follow that stems would be recovered in the absence of affixes or that the identification of these morphemes would be beneficial in all cases. Indeed, some research has suggested that affixes and stems differ in terms of their sensitivity to position (e.g., Crepaldi, Rastle, Davis, & Lupker, 2013). Moreover, the lack of facilitation from nonsuffixed words (e.g., cashew) in previous research (e.g., Beyersmann et al., 2016) suggests that the presence of a free morpheme (e.g., cash) might not be beneficial in the subsequent processing of that morpheme if the remaining part of the word (e.g., ew) is not a legal morpheme. However, some nonsuffixed items in previous experiments (see, e.g., Beyersmann et al., 2016) did contain two legal morphemes (e.g., ad + dress, beg + in, drag + on, car + rot, car + ton) and, consequently, had a pseudocompound structure, as did some of the pseudosuffixed words (e.g., leg + ion; lot + ion, miss + ion, port + ion; ion can refer to a type of atom and is an unbound morpheme, as well as having a suffix meaning). Therefore, past research does not provide an unambiguous indication of whether words with a pseudocompound structure (e.g., carpet and carrot) would benefit from the identification and access of embedded morphemes in the same way that words with a true compound structure (e.g., necklace) do.

Just as there is currently no definitive empirical evidence concerning these issues, theories of complex word processing also make very different predictions about whether compound or pseudocompound structures will be differentially influenced by the recovery of embedded morphemes depending on their representational and processing assumptions. Some theoretical approaches predict a processing benefit. For example, Crepaldi et al.’s (2010) model posits facilitatory links between units at the morpho-orthographic segmentation level (e.g., corn, er, deal) and representations in the orthographic lexicon (e.g., corn, corner, deal, dealer). Consequently, their model would predict facilitation for both pseudocompounds and compounds.

Similarly, Grainger and Beyersmann (2017, p. 300) suggested that bimorphemic nonwords (e.g., dustworth) take longer to process than nonwords consisting of a nonword and a word due to lexical activation from both embedded morphemes, which suggests that stems can be recovered even when a word does not contain affixes. If this is the case, then pseudocompounds and compounds should also show facilitation relative to an unrelated control word due to the principle of full decomposition and the two stems providing additional lexical activation, even though, in the case of pseudocompounds, the item has a false bimorphemic compound structure.

In contrast, models that adopt a competitive selection mechanism (see Andrews & Davis, 1999, for a discussion) predict that activated representations of the whole word (e.g., blackbird) and constituents (e.g., black and bird) compete and could delay identification of the whole word. Finally, approaches that posit that the system identifies all possible constituents and construct morphological structures via an interpretive composition process (Arcara, Semenza, & Bambini, 2014; Gagné & Spalding, 2009; Ji et al., 2011; Taft, 2003; Taft & Nillsen, 2013) predict that the identification of morphemes would be beneficial for compounds due to their compatibility with the actual structure, but not beneficial for pseudocompounds due to their incompatibility with the actual morphological structure (e.g., assigning car and pet to the two constituent positions in a compound structure conflicts with the actual, monomorphemic, structure of carpet and resolving this conflict increases processing time).

To investigate automatic morphological segmentation and the consequences of a false morphological structure in the context of transposed letters, we use a spelling error detection task. Although a spelling error detection task has not yet been used to examine the issue of morphological decomposition as we do in the current project, it has been used to examine other aspects of lexical access. For example, a version of this task was used by MacKay (1972, see also MacKay, 1992, for a review) to examine the influence of phonology on lexical access. We adapted this task for exploring questions about the role of orthography and morphology. In our version of the task, participants indicated, by pressing one of two keyboard keys, whether the word was spelled correctly. The spelling error detection task involves a procedure similar to a lexical-decision task, in that both require a yes/no judgment. However, they differ in that for the lexical-decision task the participants are told that some items are actual words and some are nonwords, whereas in the spelling error detection task, all items are words but some are misspelled.

The spelling error detection task has several advantages for addressing our research question. One advantage is that additional items are not needed as fillers to balance the number of yes and no responses. The spelling error condition serves this purpose, while presenting the same intended words that the participant must access in order to identify that it is misspelled. Thus, the “no” responses are as informative as the “yes” responses in that both the correctly spelled and misspelled items provide useful information for evaluating the various hypotheses. Another advantage of a spelling error detection task is that the judgment is more highly practiced, and naturalistic than is a word/nonword judgment, particularly for university students. Students routinely proofread text and check for spelling errors. Consequently, decisions about spelling more directly tap into information about word form (e.g., orthography) and do not involve metajudgments about word status, which could be based on meaning or other factors as well as form. Finally, by using letter transpositions at the morpheme boundary, we are able to examine the potential role of morphology without relying on priming. Rather than looking at the influence of the recent presentation necklace on neck, for example, or of neclkace on neck to determine whether neck becomes available during the processing of the compound necklace, we directly look at the processing of necklace or neclkace. That is, the spelling task allows us to directly examine the processing of compounds and pseudocompounds in a context where the constituents have not been recently viewed, rather than examining the consequences of a recently presented word on processing.

It is unclear from the existing literature whether legal stems in words without affixes (i.e., words containing two or more free morphemes) are always identified regardless of the actual morphological structure of the whole word and, if so, whether this segmentation differentially influences the ease of processing compound and pseudocompound words. To address these issues, the current studies investigate the effects of the presence of two adjacent free morphemes on the processing of compound (e.g., necklace) and pseudocompound (e.g., carpet) words using a spelling error detection task and transposed letters at the (pseudo-) morphemic boundary. The letter transposition introduces a spelling error into the stimuli (e.g., neclkace and capret) without changing the intended word and thus we take advantage of this aspect of the letter transposition manipulation by using a spelling error detection task to explore the role of morphemes in word recognition.

Across three experiments, we manipulate word type (compound, pseudocompound, and control words) and presence of a spelling error in order to examine the role of embedded morphemes in the context of words for which the morphemes can be correctly assigned to a compound morphological structure and for words for which the compound structure is false. In Experiment 1, we examine whether it takes less time to indicate that a compound is correctly spelled relative to a control word, and, if so, whether a letter transposition at the morpheme boundary removes this advantage by making the morphemes more difficult to detect. In Experiment 2, we examine whether the presence of morphemes influences the processing of pseudocompounds relative to a control word. In Experiment 3, compound and pseudocompound words are examined in the same experiment. For all experiments, the control words had a letter transposition at the same letter location as their frequency- and length-matched compound/pseudocompound word.

We should clarify that the primary factor that we are focusing on is whether embedded morphemes function as morphemes or not in the particular word, not on semantic transparency. The constituents of a compound can either be semantically opaque (e.g., honey and moon do not contribute to the meaning of honeymoon) or semantically transparent (e.g., blue and berry do contribute to the meaning of blueberry). In contrast, the issue of semantic transparency for a pseudocompound is not a meaningful construct. To illustrate, the pseudocompound hippie consists only of the morpheme hippie, is not formed from the morphemes hip and pie, and, consequently, the semantic transparency of hip and pie does not truly exist even though they can be thought of as being semantically opaque. Put another way, the reason why hip and pie do not contribute to the meaning of hippie is because they do not function as morphemes within that particular word. Indeed, previous research has shown that pseudocompounds show quite different processing patterns compared to fully opaque compounds even though both are semantically opaque (e.g., Gagné, Spalding, Nisbet, & Armstrong, 2018; Gagné, Spalding, & Nisbet, 2016).

Examining words for which the embedded morphemes are used productively (i.e., function as a morpheme within a particular word as is the case for compounds) and for which the embedded morphemes are not used productively (i.e., do not function as a morpheme within a particular word as is the case for pseudocompounds) allows us to explore various theoretical approaches concerning whether morphemes are automatically detected and if so whether the presence of these morphemes aids or hinders recognition of the whole word. If embedded morphemes are not recovered during the processing (i.e., if words are accessed as whole-word representations without decomposition) then compounds and pseudocompounds should not differ from frequency- and length- matched control words, and the impact of letter transpositions should be equivalent for the experimental and control words. However, if all morpho-orthographic representations are recovered and have facilitatory connections to words containing those letter sequences (e.g., Crepaldi et al., 2010; Grainger & Beyersmann, 2017), then both compounds and pseudocompounds would benefit from the recovery of morphemes. Similarly, letter transpositions which make it more difficult to identify the morphemes should decrease the extent to which the presence of the morphemes benefits word processing. However, if morpho-orthographic representations act as competitors for the whole-word representation (see Andrews & Davis, 1999 for a discussion of competitive selection), then one would expect the processing of compounds and pseudocompounds to be slowed relative to their matched control words.

Finally, by examining both pseudocompounds and compounds, we are able to disentangle a slow-down in processing due to competing morphological representations (e.g., the retrieval of pan and try during pantry, or of neck and lace during necklace) that is predicted by some existing views of word processing from the impact of morphological construction (the bimorphemic compound structure [pan] + [try] is incompatible with the pseudocompound pantry whereas the compound structure [neck] + [lace] is compatible with the actual morphemic structure of compound necklace). Competitive selection models would predict that competing morphological representations would interfere with the processing of both compounds and pseudocompounds.

However, these models do not consider the impact of the compatibility of the actual morphemic structure of the target. If the recovery of morpho-orthographic representations triggers a composition process then compounds should show a processing advantage relative to control words because morphemic composition would yield a morphological structure that is compatible with the true structure, and also the morphemes would boost activation of the compound, whereas pseudocompounds should not show this advantage because the computed compound structure (triggered by the presence of two free morphemes) would be incompatible with the actual morphemic structure of the pseudocompound (see Arcara et al., 2014; Gagné & Spalding, 2009; Ji et al., 2011; Taft, 2003; Taft & Nillsen, 2013).

Note also that an incompatible morphemic structure (such as the construction of a bimorphemic structure, hip + pie, for the pseudocompound hippie) would also result in an incompatible semantic representation. The bimorphemic compound structure would trigger a conceptual combination process (i.e., the word would be processed as a though it were a novel compound) which would result in a meaning (e.g., pie in the shape of a hip, or pie that adds to the size of your hip, etc.) that is incompatible with the usual meaning of the word. Thus, in addition to an incompatible morphemic structure producing processing issues for a pseudocompound, the resulting incompatible semantic representation could also interfere with processing in a way that has already been demonstrated for the literal meaning of semantically opaque compounds producing processing difficulties (e.g., Ji et al., 2011; Spalding & Gagné, 2014) or, more generally, for relational competition influencing processing times for compound words and novel phrases (e.g., Gagné & Spalding, 2014a). Nonetheless, recent research comparing pseudocompounds and opaque compounds (Gagné et al., 2018) has found processing difficulties due to an incompatible morphemic structure that cannot be reduced to semantic effects. Furthermore, it is possible to have a compatible morphemic structure (e.g., hogwash actually is consistent with hog + wash) but an incompatible semantic representation if the meaning is actively constructed via conceptual combination (e.g., hogwash is not a wash for pigs, but this meaning could be constructed following the identification of a bimorphemic structure). Therefore, our discussion of the issue will focus on the question of the compatibility of morphemic structures rather than on the resulting semantic representations which are afforded by such structures.

Moreover, a letter transposition could make it more difficult to recover the morphemes, which would increase the difficulty of constructing the compound structure. This increase in difficulty would attenuate the processing advantage for compound words but would (relatively) benefit the processing of pseudocompounds by reducing the interference from an incompatible morphemic structure.

Past research has shown a compound processing advantage for semantically transparent compounds in tasks such as lexical decision (Fiorentino & Poeppel, 2007; Ji et al., 2011) and typing (Gagné & Spalding, 2014c, 2016; Libben, Curtiss, & Weber, 2014) and this is consistent with research suggesting that morphemes are automatically detected (e.g., Fiorentino & Fund-Reznicek, 2009; Libben, Gibson, Yoon, & Sandra, 2003). The aim of the current experiment is to determine whether compound words show a processing advantage in a spelling error detection task, and, if so, whether this advantage is disrupted by a letter transposition at the morpheme boundary. Unlike previous work on compounds and transposed letters (e.g., Christianson, Johnson, & Rayner, 2005) that uses the item with the transposed letter as a prime (e.g., susnhine) to see the impact that manipulation has on subsequently naming a compound versus a control word (e.g., sunshine vs. sunsbine), the current experiment takes a different approach in that it directly compares the compound and transposed letter item to their frequency- and letter-matched control words. Thus, in addition to determining whether compounds undergo decomposition, we also examine the consequences of that decomposition in terms of the role that the morphemes play in word recognition (as measured by the ease with which people can decide whether the word is correctly spelled).

Manipulating the ease with which the embedded morphemes can be recovered by introducing a letter transposition at the morphemic boundary allows us to evaluate whether the compound processing advantage is due to access to morphemic information. In particular, there should be a processing advantage when the compound is correctly spelled. However, disrupting the morpheme boundary should make it more difficult to recover the morphemes (Christianson et al., 2005; Duñabeitia, Perea, & Carreiras, 2007) and, thus, should attenuate or eliminate the processing advantage for compound words. The possibility that the advantage would be attenuated rather than eliminated comes from previous research using masked priming that found that disruptions at the morpheme boundary do not entirely disrupt the recovery of morpho-orthographic units (e.g., Christianson et al., 2005; Perea & Carreiras, 2006; Rueckl & Rimzhim, 2011; Sánchez-Gutiérrez & Rastle, 2013; see also Beyersmann, McCormick, & Rastle, 2013, for a discussion of data on this issue).

The plan for this experiment and all experiments reported in the article has been reviewed for its adherence to ethical guidelines by a Research Ethics Board at the University of Alberta.

Materials

The experimental materials included 80 control and 80 fully transparent compound words (see Appendix A). Each compound word was matched with a control word in terms of SUBTLEX-US log frequency (Brysbaert & New, 2009) and letter length (within 1 letter). The control words did not have a compound structure. To ensure that there were no unintended repetition priming effects, all morphemes were unique and no words were repeated. Two adjacent letters were transposed to create a spelling error. For the compound words, the transposition was at the morpheme boundary such that the last letter of the first constituent and the first letter of the second constituent were switched (e.g., doorbell became doobrell). For the control words, the letters were transposed at a location within the word that matched the location of the switch in the matched compound word. For example, the fourth and fifth letters of particle would be transposed to make paritcle, so that it matched the position of the switch in doobrell. Thus, the compound and control pairs were matched in terms of whole-word frequency, length, and position of the relevant bigrams. The frequency of the relevant bigrams (e.g., the rb in doorbell and br in doobrell) were free to vary and their influence was controlled statistically in the analysis. Bigram frequencies were obtained from Jones and Mewhort (2004). See Table 1 for the descriptive statistics for the stimulus variables.

The stimulus lists were counterbalanced across two lists such that across the two lists every word was seen with and without the spelling error. Each list contained only one version of each stimulus (e.g., a given participant would see either doorbell or doobrell). Thus, each list had 80 compound words, of which 40 were correctly spelled and 40 had errors, and 80 control words, of which 40 were correctly spelled and 40 had errors. An additional 80 compound words and 80 control words were selected to be the fillers so that the spelling errors did not always occur in the middle of the word. Half of the fillers were spelled correctly and the others were spelled incorrectly. The spelling errors were created by switching adjacent letters at random locations within the word. Random switches within the fillers distributed the spelling errors evenly across letter positions to prevent participants from only looking at the middle of the words and skewing their reaction time (RT) over the course of the experiment. The final filler list had 80 compound words, of which 40 were correctly spelled and 40 had randomly placed errors, and 80 control words, of which 40 were correctly spelled and 40 had randomly placed errors. Each person completed 320 trials and the order of presentation was randomized for each participant.

Procedure

Each trial began with the word “Ready?” and participants pressed the spacebar to initiate the trial. Next, the stimulus appeared and remained on the screen until the participant responded. Participants responded with key ‘J’ if the word was spelled correctly and key ‘F’ if the word was spelled incorrectly.

Participants

Forty first-year psychology students at the University of Alberta participated for partial course credit. One subject was removed from the analysis due to high variability and long response times. All participants in the current experiment and in Experiments 2 and 3 were native speakers of English. Each experiment contained a unique set of participants.

The data were analyzed using linear mixed effects (LME) regression models (Rabe-Hesketh & Skrondal, 2012) in in Stata 15 with the mixed function for the response time data and the meqrlogit for the binary (correct vs. incorrect) accuracy data (for an overview of using Stata to conduct LME analysis, see http://blog.stata.com/tag/multilevel-models/). Separate analyses were conducted for the correctly spelled and misspelled items. Participant and item were entered as crossed random factors (i.e., random intercept for items and a random intercept for subjects were included in the model) due to the repeated-measure nature of the design, and word-type (compound vs. control) was entered as a fixed factor. To statistically control for the potential influence of bigram frequency, this variable also was included in the model. We report tests conducted on the estimates of the fixed effects. Because the mixed function fits the linear-mixed regression model using dummy coding for categorical variables, the coefficients (i.e., estimates) from this function correspond to simple effects rather than to main effects. Therefore, the contrast function in Stata was used to test the main effect of word-type, and the result of this test is reported as a chi-square.

Our choice of model in terms of the random effects was based on theoretical, rather than data-based, considerations about the experimental design and about the predictions of interest. This strategy is based on the observation that there is not a clear consensus in the literature on whether to use a maximal structure. Barr, Levy, Scheepers, and Tily (2013) claim that, in some cases, not using a maximal structure can be anticonservative, but others (such as Matuschek, Kliegl, Vasishth, Baayen, & Bates, 2017) point out that using a maximal structure can lead to substantially reduced power in the overall test. Including random slopes substantially increases the number of parameters and this can be problematic without a sufficiently large number of data points. Overall, it appears that the best strategy is to be guided by the theory underlying the literature, and to consider whether the maximal structure is appropriate given the data structure. In our case, there was no strong a priori theoretical rationale (based on the dominant theories in the literature) for fitting either subject-specific or item-specific slopes for the experimental factors. Nonetheless, after fitting our original models we subsequently ran models with by-item and by-subject slopes for all the fixed effects. When we compared the fit of the original model to a maximal model, there was no strong justification for using the more complex model for the majority of the analyses reported in the current article. The maximal model improved the fit only for Experiment 1, and for the analysis on the correctly spelled items in Experiment 3. More importantly, we found that in all cases the pattern of results for the fixed effects did not differ from the models we report in this article. Thus, our conclusions are unaffected by whether or not the more maximal model is used.

The descriptive statistics are shown in Table 2. The response times were log-transformed to reduce skewness in the residuals. Only trials with the correct response were included in the response time analysis and responses less than 350 (n = 4) were removed as outliers. Following the transformation and the removal of these outliers, visual inspection of the residual plot showed no obvious violations of the normality assumption or of homoscedasticity.

Response time analysis

When the words were spelled correctly, responses to the compounds were faster than were responses to the control words, z = 3.92, p < .0001. This is consistent with past research showing that compound words are processed more quickly than noncompound control words of similar length and frequency (e.g., Ji et al., 2011). However, as predicted, this advantage for compound words disappeared when the words contained a letter transposition (i.e., when the morpheme boundary is disrupted), z = 1.18, p = .24. In terms of the control variable, response time for correctly spelled items was unaffected by bigram frequency, b = .003, SE = .003, z = 1.27, p = .20, but the response time for incorrectly spelled items increased as bigram frequency increased, b = .004, SE = .002, z = 2.22, p = .03.

Accuracy analysis

For correctly spelled words, accuracy was higher for compounds than for their matched controls, but this effect was not significant, z = −1.92, p = .055. For incorrectly spelled words, accuracy was higher for compounds than for their matched controls, z = −2.59, p = .01. In terms of the control variable, accuracy for both correctly spelled items and incorrectly spelled items was unaffected by bigram frequency, b = −.04, SE = .08, z < 1, and b = −.02, SE = .04, z < 1, respectively.

Alternative analysis

Analyzing the misspelled and correctly spelled words separately is consistent with the strategy used in analyzing lexical decision data. Namely, in lexical decision tasks, researchers typically do not compare the response times to word and nonword responses because, by hypothesis, the word and nonword responses must involve very different processes. In particular, the idea is that the “word” response occurs when the incoming letter string accesses the word, but the “nonword” response can only occur when the person reaches some separate criterion that the letter string has mismatched all the words in the lexicon (because the question is not whether it matches a particular word, but whether it is a word at all). The spelling error detection task, on the other hand, does not have this kind of closed process versus open process associated with the two outcomes. In particular, previous research (e.g., Perea & Carreiras, 2006; Rueckl & Rimzhim, 2011) shows clearly that having transposed letters does not stop the participant from accessing the intended word. Thus, the “correctly spelled” response occurs when the spelling of the letter string matches the stored orthography of the intended word, and the “incorrectly spelled” response occurs when the letter string mismatches that stored orthography. Thus, the two responses in the current spelling error detection task do not correspond to large processing differences in the way that they do in lexical decision.

Therefore, another analysis strategy is to analyze the misspelled and correctly spelled conditions within the same analysis. Thus, we fit LME regression models using participant and item as crossed random factors, and word-type (compound vs. control) and spelling (no error vs. spelling error) as fixed factors. Bigram frequency was entered as a control variable.

The pattern of data are identical for that observed in the separate analysis strategy. The response time analysis showed an interaction between word-type and spelling, χ2(1) = 15.94, p < .0001. Therefore, we followed up on this interaction by examining the simple effect of word-type within each level of spelling. When the words were spelled correctly, responses to the compounds were faster than were responses to the control words, z = 5.14, p < .0001. However, as predicted, this advantage for compound words disappeared when the words contained a letter transposition (i.e., when the morpheme boundary was disrupted), z = 1.38, p = .17. Another way to follow up on this interaction is to examine the impact of spelling for each word-type. Both compound words, z = −13.39, p < .0001, and control words, z = −7.29, p < .0001, were slowed by letter transposition, but control words were less affected by the transposition (as indicated by the interaction).

The accuracy analysis did not show an interaction between word-type and spelling, χ2(1) < 1. In terms of the main effects, compound words were responded to more accurately than the control words, χ2(1) = 12.31, p < .0005, and accuracy was higher for correctly spelled words than for incorrectly spelled words, χ2(1) = 83.67, p < .0001.

In sum, the results suggest that it is easier to process compound words compared to the matched control words even when bigram frequency was statistically controlled, and that in the response time data this advantage is eliminated when there is a letter transposition at the morpheme boundary. Recall that compounds and their controls were matched pairwise in terms of word frequency, and, consequently, the observed processing advantage cannot be due to compounds being more familiar. Thus, it appears that access to the constituents aids access to the compound and this leads to a processing benefit. A disruption at the morpheme boundary interferes with morphemic decomposition and limits the ability of the constituent representations to aid access of the compound in terms of speed of processing, but still benefited processing in terms of the accuracy data. Finally, the results clearly show that the spelling error detection task is highly sensitive to the morphological structure of the word, and hence is a suitable task for investigating the role of morphology in processing.

We suggest that the compound advantage observed in Experiment 1 is a consequence of morphological decomposition and the involvement of morphological constituents during the recognition of the compound. To further explore this possibility and to gain insight into the processing of items that contain embedded morphemes but do not have a compound structure, we investigated the processing of pseudocompounds. Pseudocompounds (e.g., carpet and lotion) have the appearance of compound words, but in fact do not have a compound morphemic structure (e.g., carpet looks like a compound composed of car and pet).

The theoretical questions of whether embedded morphemes in a pseudocompound are automatically detected, and whether the presence of these morpho-orthographic representations helps or hinders recognition of the pseudocompound has not yet been fully explored. The prior literature suggests three general possibilities. First, if embedded morphemes are not recovered during the processing (i.e., if words are accessed as whole-word representations without decomposition) then the presence of embedded morphemes is irrelevant and pseudocompounds should not differ in ease of processing from length- and frequency-matched control words. Also, the impact of a letter transposition should be equivalent for pseudocompound words and their matched controls (e.g., each would be disadvantaged to the same degree). Second, if all morpho-orthographic representations are recovered and have facilitatory connections to words containing those letter sequences regardless of a word’s true morphological structure (e.g., rose is connected to rosebud and rose and car is connected to carpet and car), then the recovery of the pseudomorphemes would aid participants’ ability to determine whether the word was correctly spelled. Similarly, letter transpositions, which make it more difficult to identify the morphemes, should decrease the extent to which the presence of the morphemes benefits word processing by slowing the access of the morpheme representations. Finally, if morpho-orthographic representations are recovered and, perhaps, used to construct a morphological structure then pseudocompounds would be more difficult to process relative to length- and frequency-matched control words because the constructed morphological structure is incompatible with the actual structure of the word. In addition, the recovered morphemes might serve as competitors for the pseudocompound. In either case, unlike compounds, pseudocompounds would be hindered by the recovery of embedded morphemes.

Materials

The experimental items consisted of 80 control words and 80 pseudocompound words (see Appendix B). The pseudocompound words could be parsed into two free morphemes but, unlike compound words, these morphemes do not function as such in the pseudocompound (e.g., lotion contains the English morphemes lot and ion, but is mono-morphemic). Each pseudocompound word was matched with a control word in terms of SUBTLEX-US log frequency and letter length (within one letter). To create a spelling error in the pseudocompound words, the adjacent letters at the embedded-morpheme boundary were switched (e.g., carpet became capret). The control words were created in the same manner as in Experiment 1. The letters were switched at the same location within the word as in their matched pseudocompound words. Thus, the pseudocompounds and their control word were matched in terms of word frequency, length, and position of the transposed letters. Bigram frequency of the relevant bigrams was obtained from Jones and Mewhort (2004) for inclusion in the analysis as a control variable. See Table 3 for the descriptive statistics for the stimulus variables.

The stimuli were counterbalanced so that each list included only one version (i.e., correctly spelled or misspelled) of each word. In total, each list had 80 control words, 40 correctly spelled and 40 with an error, and 80 pseudocompound words, 40 correctly spelled and 40 with an error. In addition to the experimental stimuli, each participant saw a set of 80 control words and 80 pseudocompound words as fillers. For half of the filler words, a spelling error was created in a random location, using the same steps as in Experiment 1, to distribute the spelling errors evenly throughout the word. All the other words were spelled correctly. The filler list had 80 pseudocompound words, of which 40 were correctly spelled and 40 had an error, and 80 control words, of which 40 were correctly spelled and 40 had an error.

Procedure

The procedure was identical to Experiment 1.

Participants

Fifty first-year psychology students at the University of Alberta participated for partial course credit.

The descriptive statistics are shown in Table 4. The data were analyzed using separate linear mixed effects regression models for the response time and accuracy data (see Experiment 1 for details). Separate models were fit for the correctly spelled and misspelled items. Word-type was entered as a fixed effect and item and subjects were entered as crossed random effects. Bigram frequency was entered as a control variable. Response times were log transformed. Only trials with the correct response were included in the RT analysis and responses less than 150 (n = 4) were removed as outliers. One item in the control condition was removed because it had a pseudocompound structure.

Response time analysis

In the correctly spelled condition, pseudocompounds did not take longer than their matched controls, z < 1. In the misspelled condition, pseudocompounds took longer to process than did their matched controls, z = 2.45, p = .01. Bigram frequency did not influence response times for the correctly spelled items, b = −.002, SE = .004, z < 1, or for misspelled items, b = .003, SE = .003, z < 1.

Accuracy analysis

The correctly spelled pseudocompounds did not differ from their controls, z = −1.35, p = .18. In the misspelled condition, pseudocompounds did not differ in terms of accuracy from their matched controls, z = −1.47, p = .14. Bigram frequency did not influence accuracy for either the correctly spelled items, b = −.03, SE = .12, z < 1, or for misspelled items, b = .01, SE = .05, z < 1.

Alternative analysis

As in Experiment 1, additional analyses were conducted in which the misspelled and correctly spelled items were analyzed within the same model. In the response time analysis, word-type and spelling interacted, χ2(1) = 5.59, p = .018. In the misspelled condition, pseudocompounds took longer to process than did their matched controls, z = 9.78, p = .002. However, in the correctly spelled condition, pseudocompounds did not differ from their matched controls, z < 1. Another way to follow up on this interaction is to examine the impact of spelling for each word-type. Both pseudocompound words, z = −7.09, p < .0001, and control words, z = −3.94, p < .0001, were slowed by letter transposition, but control words were less affected by the transposition (as indicated by the interaction).

In the accuracy analysis, there was no interaction between word-type and spelling, χ2(1) = 1.29, p = .26. There was a main effect of word-type in that accuracy was lower for pseudocompounds than for the control words, χ2(1) = 5.43, p < .02, but no main effect of spelling, χ2(1) = 2.64, p = .10.

To determine whether pronunciation match or mismatch between the pseudomorpheme and the pseudocompound influenced spelling detection, we conducted an additional analysis using the correctly spelled pseudocompounds in which we included whether the pronunciation of the first pseudomorpheme (e.g., son) was maintained in the pseudocompound (e.g., sonnet), along with the bigram frequency as a control variable. Bigram frequency did not influence response time, z < 1 nor did it influence accuracy, z < 1. Response times for the correctly spelled pseudocompounds were influenced by whether the pronunciation of the pseudomorphemes matched the pronunciation of the pseudocompound, χ2(1) = 6.93, p = .009, in that it took less time to indicate that the pseudocompound was correctly spelled when the pseudomorpheme’s pronunciation was retained in the whole word than when the pseudomorpheme’s pronunciation differed from the related part of the whole word. This finding suggests that the first constituent was being extracted during the processing of the pseudocompound, because otherwise pronunciation of that pseudoconstituent should have no influence on the processing of the pseudocompound. That is, the pronunciation of son could only be relevant if it was being accessed during the processing of sonnet. Pronunciation match did not predict accuracy, χ2(1) = 2.80, p = .09, although there was a tendency for performance to be more accurate when the pronunciation matched than when the pronunciation mismatched.

In sum, the differences in processing for the pseudocompounds and their matched controls indicate that the system attempts to recover morphemes, even if they do not play a morphological role in the target word. The presence of embedded morphemes increased the processing difficulty of misspelled pseudocompounds relative to their matched control words. In addition, the effect of the pronunciation match between the constituent and the pseudocompound also suggests some access to the constituents. Importantly, the pattern of data observed in the current experiment for pseudocompounds is opposite to what was observed in Experiment 1 for compounds even though both the compounds and pseudocompounds in these experiments contain two embedded morphemes, which indicates that the true morphemic structure of the word influences the impact of the embedded morphemes.

Experiments 1 and 2 revealed that the ease of processing pseudocompounds and compounds differed from their control words. Compounds were easier to process than the length- and frequency-matched control words, whereas pseudocompounds were more difficult to process than their control words when misspelled. The compound advantage effect was attenuated in the response time data for compounds when a letter transposition was introduced into the word at the morphemic boundary. In contrast, pseudocompounds were more difficult to process relative to the control word when they contained a letter transposition.

In this experiment, we examine both types of words within the same set of participants to determine whether the different patterns of results for pseudocompounds and compounds observed in the previous experiments were due to participant differences rather than due to word-type (i.e., compound vs. pseudocompound) differences. In addition, in Experiments 1 and 2 some of the control words contained plurals and this might have aided their processing. Thus, in the current experiment, we used only singular words. In the analysis, we also examine the possible influence of additional variables relating to the legality of the letter sequences and syllable structure.

Materials

We created the stimulus list by combining the word lists from Experiment 1 and 2 (see Appendix C). However, plural words and two duplicated words (which were used as a control word in both Experiment 1 and 2) were replaced.

The experimental items consisted of 80 fully transparent compound words (40 spelled correctly and 40 misspelled), 80 pseudocompound words (40 spelled correctly and 40 misspelled), and 160 control words (80 spelled correctly and 80 misspelled). The filler items mimicked the structure of the experimental set except that for the misspelled items the adjacent letters were switched at random positions within the words to create the spelling errors. The filler items consisted of 80 compound words (40 spelled correctly and 40 misspelled), 80 pseudocompound words (40 spelled correctly and 40 misspelled), and 160 control words (80 spelled correctly and 80 misspelled).

Procedure

The procedure was identical to Experiments 1 and 2.

Participants

Sixty first-year psychology students at the University of Alberta participated for partial course credit. The data from six participants were not included in the analysis: one participant talked during the experiment, one participant had extremely fast (less than 100 ms) responses, and four participants had low (below 60%) accuracy rates.

The descriptive statistics are shown in Table 5. Only trials with the correct response were included in the response time analysis. A log transformation was used to correct for skewness in the residuals. Visual inspect of the residuals indicated that responses less than 380 ms or greater than 10 s were outliers and these were removed before fitting the final model.

As in Experiments 1 and 2, separate linear mixed effects regression models were fit for the response time and accuracy data (see Experiment 1 for details). Misspelled and correctly spelled items were analyzed in separate models. In these models, set (i.e., compound set vs. pseudocompound), and word-type (i.e., pseudocompound or compound = 0 vs. control = 1) were entered as fixed effects and item and subjects were entered as crossed random effects. Bigram frequency was entered as a control variable.

Response time analysis

For the correctly spelled items, there was an interaction between set and word-type, χ2(1) = 19.10, p < .0001. The simple effects analyses revealed that correctly spelled compounds were more quickly processed than their matched controls, z = 6.02, p < .0001, but pseudocompounds did not differ from their matched controls, z < 1. The control variable, bigram frequency, was not a valid predictor, b = .003, SE = .003, z = 1.27, p = .20. The misspelled items also revealed an interaction between set and word-type, χ2(1) = 9.48, p = .002. The simple effects analyses revealed that compounds did not differ from their matched controls, z = 1.01, p = .31, but pseudocompounds took longer to process than their matched controls, z = −3.32, p = .001. The control variable, bigram frequency, was not a valid predictor, b = .006, SE = .004, z = 1.66, p = .10.

Accuracy analysis

For the correctly spelled items, set and word-type interacted, χ2(1) = 19.54, p < .0001. The simple effects analyses revealed that compounds had higher accuracy than their matched controls, z = −5.25, p < .0001, but pseudocompounds did not differ from their matched controls, z < 1. The control variable, bigram frequency, was not a valid predictor, b = .003, SE = .002, z = 1.77, p = .08.

The misspelled items did not show an interaction between the experimental variables: set and word-type, χ2(1) < 1. There was a main effect of set, χ2(1) = 6.58, p = .01, in that the compound set had lower accuracy than the pseudocompound set, but not a main effect of word-type (i.e., control vs. noncontrol), χ2(1) = 1.00, p = .32.

Alternative analysis

As in Experiments 1 and 2, additional analyses were conducted in which the misspelled and correctly spelled items were analyses within the same model. Set (i.e., compound set vs. pseudocompound), control (i.e., pseudocompound or compound = 0 vs. control = 1), and spelling (correctly spelled vs. misspelled) were entered as fixed effects and item and subjects were entered as crossed random effects. Bigram frequency was entered as a control variable. The response time analysis revealed a three-way interaction between the experimental variables: set, word-type and spelling, χ2(1) = 3.89, p < .049.

We followed up on this three-way interaction by examining whether the influence of word-type differed for the compound set and pseudocompound set within each level of spelling and found that set and word-type interacted for both the correctly spelled z = 4.91, p < .0001 and misspelled items, z = 3.33, p = .001. The analyses of the simple effects revealed that, as in Experiments 1 and 2, compounds and pseudocompounds differed in terms of whether they showed a processing advantage or disadvantage relative to their matched controls. For correctly spelled items, compounds were faster than their matched control, z = 7.99, p < .0001, whereas pseudocompounds did not differ from their matched control, z = 1.20, p = .23. For the misspelled items, pseudocompounds were slower than their matched control, z = −3.68, p = .0001, whereas compounds did not differ from their matched control, z = 1.03, p = .30. This pattern is consistent with what we observed in the separate analyses.

A second way to break down the three-way interaction is to examine the Set × Spelling interaction for each word-type. Only the compounds and pseudocompounds were influenced by letter transpositions. Set and spelling interacted for the experimental items (i.e., the compounds and the pseudocompounds), z = −3.33 p < .0001. The follow-up analysis of the interaction indicated that the compounds were faster than pseudocompounds when the words were correctly spelled, z = −2.22, p = .03 but not when the words were spelled incorrectly, z < 1. Set and spelling did not interact for the control items, z < 1. For these words, there was a main effect of set, z = 29.89, p < .0001, but not of spelling, z < 1.

Another way to examine the three-way interaction is to consider whether spelling differentially influences the impact of word-type within each set. Spelling and word-type interacted for both the compound set, z = 8.65, p < .0001 and the pseudocompound set, z = 5.84, p < .0001. We followed up on these interactions by analyzing the simple effects. For the compound set, the misspelled compounds were slower than the correctly spelled compounds, z = −12.74, p < .0001, but misspelled control words did not differ from correctly spelled control words, z < 1. For the pseudocompound set, the misspelled pseudocompound were slower than the correctly spelled pseudocompound, z = −7.51, p < .0001, but misspelled control words did not differ from correctly spelled control words, z < 1.

In the accuracy analysis, we also observed a three-way interaction between set, word-type, and spelling, χ2(1) = 45.53, p < .0001. The control variable, bigram frequency, was negatively associated with response time, b = −.047, SE = .02, z = −2.55, p = .01.

Because of the three-way interaction, we examined the two-way interaction of set and word-type for each level of spelling. Set and word-type interacted for the correctly spelled items, z = −6.65, p < .0001. The analyses of the simple effects revealed that for correctly spelled items, pseudocompounds did not differ from their matched control, z = 1.43, p = .15, whereas compounds had higher accuracy than their matched controls, z = −7.54, p < .0001. Set and word-type did not interaction for the misspelled items, z < 1. For these items, there was main effect of set in that items in the compound set had higher accuracy than did items in the pseudocompound set, z = −3.04, p = .002. There was no effect of word-type, z < 1.

We also examined the Set × Spelling interaction for each word-type. Set and spelling interacted for the experimental items (i.e., the compounds and the pseudocompounds), z = 9.98, p < .0001, but not for the control words, z = 1.05 p = .30. The follow-up analysis of the interaction indicated that the compounds had higher accuracy than did pseudocompounds when the words were correctly spelled, z = 7.14, p < .0001, but not when the words were misspelled, z = −1.59, p = .11. Set and spelling did not interact for the control items, z = 1.05 p = .30. The analyses of the main effects revealed that control words in the compound set had lower accuracy than did control words in the pseudocompound set, z = −2.59, p = .009. Also, correctly spelled control words had lower accuracy than did their misspelled counterpart, z = −4.36, p < .0001.

Examination of other variables that might influence processing

The control words were length- and frequency-matched to the pseudocompounds and compound words because, based on past research, these variables are likely to influence lexical access. In addition, the letter transpositions occurred at the same letter position for the experimental item and its control. However, other variables were free to vary. In this section, we examine whether our results of interest are impacted by the inclusion of variables related to orthography and phonology. In particular, we investigate the impact of (a) whether the transposed letters involved a vowel, (b) the bigram frequencies of all letter pairs impacted by the transposition, (c) whether the transposition affected the syllable structure around the morpheme boundary, and (d) the number of syllables. In this section, we will describe each of these variables before reporting an analysis that includes these additional control variables.

We did not specifically match the items in terms of whether the transposed letters equally involved vowels and consonants. For example, two consonants (i.e., rb) are exchanged to produce the misspelled version of doorbell, whereas a vowel and a consonant (i.e., ti), are exchanged to produce the misspelled version of its control word (i.e., particle). In this example, the resulting bigrams are still legal (e.g., br and it), and in the case of the compound is actually a more frequent sequence which, if spelling error detection relied on orthographic information rather than on morphological extraction, would predict a benefit for compounds over their controls. To address this issue, we coded whether the letter transposition involved a vowel or not. For the pseudocompound set, 66 of 80 of the control words involved a vowel transposition whereas 34 of 80 of the pseudocompounds involved a vowel transposition. For the compound set, 60 of 80 of the control words involved a vowel transposition whereas 14 of 80 of the compounds involved a vowel transposition. Thus, for both sets, the control word was more likely to have a vowel transposition than the experimental word and, consequently, our finding that pseudocompounds and compounds behave differently cannot be attributed to differences in the number of items that involve vowel transpositions.

To further address the issue of legality, we also considered the frequency of the other two pairs of letters that are influenced by the letter transposition. To illustrate, transposing rb in doorbell produces doobrell. In the original set of models, we had entered the frequency of the letters relevant to the transposition (e.g., rb for doorbell and br as for doobrell) and found that bigram frequency at the boundary was not a valid predictor for either the correctly spelled or misspelled items. We extended our investigation by also considering the bigram frequency of the letters prior to the boundary (e.g., ob for doobrell) and the bigram frequency of the letters after the boundary (e.g., re).

In addition to orthographic factors, we also considered whether syllable structure influenced the pattern of results that we had initially observed. In particular, we considered the number of syllables, and whether the letter transpositions fell at a syllable boundary. We examined this possibility by focusing on the misspelled condition and coding whether the bigram switch occurred at the boundary (e.g., p and k were transposed in pump-kin to form pumkpin; m and a were transposed in cin-na-mon to form cinnmaon) or not (e.g., s and a were transposed in pleas-ure to form plesaure; i and l were transposed in ap-pli-ance to form appilance). Syllable structure was obtained using the WordData function in Mathematica (see Appendix D for a list of items and their syllable structure) and was compared to the structure reported in the American Heritage dictionary. The structures were identical in both sources except for six items (all control words) and these differences reflected pronunciation variations (e.g., ome-lette vs. om-e-lette, ex-pend-i-ture vs. ex-pen-di-ture, and ref-er-ee vs. ref-e-ree). We used the structures from Mathematica in our analysis, but also conducted an analysis using the syllable structures from the American Heritage dictionary and found the same pattern. All of the compounds (80 of 80) involved a transposition at the syllable boundary (as would be expected given that the transposition involved the morpheme boundary), and 43 of 80 of the pseudocompounds occurred at the syllable boundary. The matched controls were nearly equivalent across the two sets with 31 of 80 in the pseudoset and 36 of 80 in the compound set involving a syllable boundary.

We conducted further analyses to determine whether the pattern of results still held after these potential confounds were accounted for. Some variables, such as bigram frequency and number of syllables, apply to both correctly and misspelled items, whereas the variables relating to transpositions apply only to the misspelled items. Thus, we conducted separate models for these two situations.

For the misspelled items, we used the variables concerning vowel transpositions, legality of the letter strings, and syllable structure. Specifically, we fit a model that included vowel transposition (yes/no), before-bigram frequency, after-bigram frequency, and boundary bigram frequency, syllable boundary (yes/no), and number of syllables along with our two experimental variables of interest (set and word-type). Item and subjects were included as cross-random effects. Our results confirm what we observed in the original analysis, set and word-type interacted, χ2(1) = 4.54, p = .03. The simple effects revealed that pseudocompounds were slower than their matched control, z = −3.12, p = .0002, whereas compounds did not differ from their matched control, z < 1. As for the control variables, neither afterboundary bigram frequency (z = 1.57, p = .12) nor syllable boundary (z <1) were successful predictors. However, vowel transposition (z = −4.17, p < .0001) before-boundary bigram frequency (z = 3.01, p = .003), boundary bigram frequency (z = 2.71, p = .007), and number of syllables (z = 8.58, p < .0001) did predict response time.

In terms of the accuracy analysis, as in the original analysis, misspelled items did not show an interaction between the experimental variables: set and word-type, χ2(1) < 1. There was a main effect of set, χ2(1) = 7.32, p = .007, in that the compound set had lower accuracy than the pseudocompound set. The main effect of word-type (i.e., control vs. noncontrol) did not reach statistical significance, χ2(1) = 3.68, p = .06.

For the correctly spelled items, the relevant variables are the bigram frequency variables and the number of syllables. We fit a model that included before-boundary bigram frequency, after-boundary bigram frequency, bigram frequency, and number of syllables along with our two experimental variables of interest (set and word-type). Item and subjects were included as cross-random effects.

As in the original analysis, set and word-type interacted, χ2(1) = 13.20, p = .0003. Tests of the simple effects indicated that compounds were easier to process than their matched controls, z = 4.05, p < .0001, whereas pseudocompounds did not differ from their controls, z < 1. In terms of the control variables, only the number of syllables (z = 6.49, p < .0001) was a valid predictor of response time. None of the bigram frequency variables were valid predictors, all zs < 1.

In the accuracy analysis, set and word-type interacted, χ2(1) = 16.10, p < .0001. The simple effects analyses revealed that compounds had higher accuracy than their matched controls, z = −4.55, p < .0001, but pseudocompounds did not differ from their matched controls, z < 1. None of the control variables were valid predictors: boundary bigram frequency (z = 1.24, p = .22), after-boundary bigram frequency (z = −1.43, p = .15), before-boundary bigram frequency (z < 1), and number of syllables (z < 1).

Summary

We see a clear replication of the patterns of Experiments 1 and 2 in that pseudocompounds and compounds produced opposite effects. Relative to their control words, the compounds show a processing advantage (in both response time and accuracy) when correctly spelled whereas the pseudocompounds show a processing disadvantage (in terms of response time) when incorrectly spelled. These effects hold even after accounting for the presence of vowel transpositions, legality of the letter strings (as reflected by bigram frequencies), and two aspects of syllable structure.

The current research investigated the role of morphemic processing in compound and pseudocompound words using a spelling error detection task to examine the impact of embedded morphemes. Across the three experiments, we find consistent results. First, we find clear evidence that compounds and pseudocompounds both differ from matched control words, strongly suggesting that some form of access of the embedded morphemes occurs. Second, we find that the effects of this access differ markedly for compounds and pseudocompounds; the access of embedded morphemes appears to decrease the difficulty of processing compound words but did not affect the difficulty of processing pseudocompound words, relative to matched control words (e.g., people were faster and more accurate when responding to necklace than to its matched control). A letter transposition at the morpheme boundary removed the processing advantage in terms of speed of processing for compound words relative to their matched controls, but created a processing disadvantage for pseudocompounds (e.g., neclkace did not differ from its matched control whereas patnry was slower than its matched control). In no case did we observe facilitation for the pseudocompounds, and, the fact that we observed these deleterious effects for pseudocompounds strongly suggests that the access of embedded morphemes is obligatory, even when doing so incurs a processing cost.

These data are useful for evaluating various ideas about the role of morphology in lexical access. As described in the Introduction, different theoretical approaches make quite different predictions concerning whether the embedded morphemes should be accessed, and what the results of that access should be on the overall processing of the word. The current data rule out three general approaches. First, any approach that assumes no access of the embedded morphemes, or that assumes access of morphemes only for morphologically complex words (such as compounds), is not consistent with the current set of findings. Second, any approach that assumes simple facilitatory links for all embedded morphemes regardless of the morphological structure of the whole word is also inconsistent. Third, any approach that assumes facilitatory links for compounds and no connections for pseudocompounds does not explain the current results. Contrary to these expectations, the current data provides strong evidence of the (attempted) use of the embedded morphemes in both compounds and pseudocompounds, and that this attempted use is helpful for compounds when correctly spelled but harmful for pseudocompounds when misspelled. This aspect of the data is consistent with findings reported by Amenta, Marelli, and Crepaldi (2015) which show that the morpheme detection is beneficial for both morphological complex (e.g., DEALER) and pseudocomplex words (e.g., CORNER) but produces inhibition at the semantic level for pseudocomplex words (for a similar result from both eye-tracking and lexical decision, see Marelli, Amenta, Morone, & Crepaldi, 2013).

The data support theoretical approaches that posit obligatory decomposition and construction. Although the compound data is consistent with the notion that morphemes are accessed and have facilitatory connections to the whole word, the pseudocompound data suggest that embedded morphemes are accessed (even when they do not function as morphemes in the whole word) and that some form of competition or inhibition occurs when the embedded morphemes in pseudocompounds are accessed. Thus, the pattern of data is relatively consistent with previous suggestions that embedded morphemes are accessed (e.g., Arcara et al., 2014; Gagné & Spalding, 2009; Libben, 1994; Taft & Forster, 1975, 1976; Taft & Nillsen, 2013), and perhaps that the impact of retrieving these morphemes is influenced by an obligatory construction process in which the presence of two (or more) unbound morphemes trigger the construction of compound structure (Gagné & Spalding, 2014b, 2014c, 2016; Ji et al., 2011; Spalding & Gagné, 2011). Whether the presence of morphemes will help or hinder processing differs for compounds and pseudocompounds because the embedded morphemes match the actual morphological structure for compounds, but mismatch for pseudocompounds. For example, the recovery of neck and lace support the construction of [neck] + [lace], and the recovery of car and pet supports the construction of [car] + [pet]. However, in the former, the construction is consistent with the actual morphological structure of the stimulus, whereas in the latter, it is inconsistent. Consequently, the embedded morphemes are helpful in processing the compounds but unhelpful in processing the pseudocompounds and this is consistent with the research showing that subsequent processing of embedded morphemes is faster when the prime was a compound, but slower when the prime was a pseudocompound (Gagné et al., 2018).

The effect of letter transpositions at the morpheme boundary provides additional insight into the nature of processing. For compounds, letter transpositions at the boundary remove the processing advantage in response time; necklace took less time to process than its matched control whereas neclkace did not. This pattern is consistent with the suggestion that the advantage arises due to the system being able to use the morphemes either to access the compound or to form a constructed morphological structure that is consistent with the actual structure of the word. Disrupting the ease of recovering the embedded morphemes (by the letter transposition) removes the advantage in terms of speed of processing.

However, the impact of letter transpositions on the pseudocompounds is less straightforward in that disrupting the pseudomorpheme boundary resulted in a processing disadvantage. These data are inconsistent with the most straightforward prediction concerning the role of recovered morphemes in which the recovery of embedded morphemes during the processing of the pseudocompound interferes with the processing of the word, but that this interference is reduced when the pseudoboundary was disrupted. Instead, our data suggest that the effect of the letter transposition is multifaceted and that we must separately consider the impact of both orthographic and morphemic processing. In particular, it is useful to move away from the assumption that the access of morphemes might be harmful to the processing of pseudocompounds, and instead consider the possibility that these two aspects of processing might produce opposite effects - namely, there might be facilitation due to orthographic identification but a slow-down due to the presence of morphemes.

Thus far, we have only considered the effects of the accessed embedded morphemes matching or mismatching the required structure of the word. However, as described in the introduction, many theories have separated orthographic and morphological levels of representation and processing (e.g., Crepaldi et al., 2010; Rastle, Davis, & New, 2004). If we assume that there are two levels of representation, we might ask what kinds of effects we should expect from the compounds and pseudocompounds, and how those effects should be affected by letter transpositions. In particular, we will consider the possible outcomes that could arise from the operation of orthographic processing and of morphological processing. We can make a few reasonable assumptions about how the system functions. First, the orthographic match involved in having embedded morphemes should aid the orthographic processing of the words (i.e., both compound and pseudocompound) because the orthographic units should help to activate the overlapping letters. Thus, apple in applesauce helps to activate the letters a-p-p-l-e and car in carpet helps activate the letters c-a-r. This then helps to activate the whole words (e.g., a-p-p-l-e-s-a-u-c-e activates applesauce and c-a-r-p-e-t activates carpet), relative to the control words, because, in essence, the letter units get some activation both from the whole words and from the embedded morphemes (e.g., the letter detector for “c” in the first position gets some feedback activation from both car and carpet, whereas a letter in the control word would only get feedback activation from the whole word). To illustrate further, the existence of the orthographic form CAR which both receives and sends activation to the letters c, a, and r, provides an additional place for the letter detectors to accumulate activation which also benefits the activation of the orthographic form CARPET. Thus, the orthographic component should produce a benefit in terms of accessing the whole word for both compounds and pseudocompounds relative to the matched control words. However, one must also take into account the consequences of activating associated representations at other levels which may or may not offset this benefit. For example, the orthographic forms CAR and PET boost the activation of the CAR and PET at the lemma level and trigger the construction of a compound morphological structure (e.g., CAR + PET). Therefore, our second assumption is that accessing embedded morphemes should affect morphological processing, making processing easier for compounds due to the morphological structure being consistent with the actual structure (e.g., applesauce is composed of the morphemes apple + sauce) and harder for pseudocompounds due to the morphological structure being inconsistent with the actual structure (e.g., carpet is not composed of car + pet).

The net effect of these two processes differs for compounds and pseudocompounds and depends the relative balance of the facilitation and inhibition. We will first consider correctly spelled items before discussing the misspelled items. For compounds, the orthographic process is beneficial, and the morphological process is also beneficial so the net effect should be facilitation. This was confirmed in our experiments. For pseudocompounds the orthographic process is beneficial, but the morphemic process is not. Our results indicate that the inhibitory effect of the morphological process offset the facilitatory effect of the orthographic process such that the net effect was that pseudocompounds did not differ from their matched controls.

These assumptions about the two processes are consistent with data (Gagné et al., 2018) showing that the processing of a target word in a lexical-decision task benefited from orthographic overlap between the prime and target only when when the target functioned as a morpheme in the prime (e.g., hogwash-hog or heater-heat) or when the target was a nonword (e.g., giraffe-gira). However, facilitation due to orthographic overlap between the target and prime was not observed when the prime was a pseudocompound (e.g., sonnet-son) or had no morphological relationship to the target (e.g., pupil-pup).

This explanation can also account for the results for the misspelled items. Previous findings (e.g., Gomez, Ratcliff, & Perea, 2008; Norris, Kinoshita, & van Casteren, 2010; Perea & Lupker, 2004) show that letter transposition is imprecisely encoded and that primes with transposed letters can aid the processing of the subsequent target (e.g., obey is aided by exposure to oeby) which suggests that even with mismatches in letter position, recovery of the orthographic and lemma representations is still possible (see also Chambers, 1979; O’Connor & Forster, 1981; Taft & Nillsen, 2013). Our results suggest that letter transpositions change the relative impact of orthographic and morphemic processing on the net effect of these two processes. For compounds, the letter transposition makes it more difficult to access embedded morphemes. The net result would be increased processing difficulty for the compound words which would either attenuate or, as observed in the current data, remove the overall advantage. For the pseudocompounds, the increased difficulty in orthographic processing when the item is misspelled, would either remove or reduce the processing advantage for the orthographic processing which would create an overall decrement for the pseudocompounds because the detrimental effect of the morphological process would no longer be offset by the the facilitatory effect (relative to the control words) of the orthographic process. The benefit at the orthographic level due to the embedded morphemes would be reduced because activation of the forms CAR, PET, and CARPET would be reduced due to mismatches in letter position. Consequently, the activation of CARPET would be directly impacted, and also indirectly impacted due to the forms CAR and PET being impacted which would reduce activation from the letter detectors. However, even with reduced activation of these forms, the lemma and morphological units would still be accessible (just as boey led to the activation of obey at the orthographic and lemma levels) which allows for morphemic construction. Note that if the whole-word representation (e.g., carpet) was the only representation available to the system (i.e., if embedded morphemes were not recovered), then pseudocompounds would not differ from their matched controls. However, this is not what our data indicate. Given that the primary difference between the pseudocompounds and the control words is that the pseudocompounds can be parsed into two morphemes, it appears that orthographic units that correspond to existing English morphemes (e.g., car and pet when either carpet or capret are presented) also are retrieved.

One might wonder—if embedded morphemes can still be activated when the word is misspelled, why does their access hurt pseudocompounds but fail to benefit compounds relative to the matched control words? To understand this apparent paradox, it is useful to remember that that pseudocompounds and compounds differ in terms of whether responses are more strongly influenced by the embedded morpheme representations (as is the case for compounds) or by the whole-word representation (as is the case for pseudocompounds). Also, for compounds, the availability of the embedded morphemes is beneficial due to facilitatory links at the lemma level and also due to supporting the construction of a consistent morphemic structure. In contrast, for pseudocompounds the availability of the embedded morphemes is detrimental due to inhibitory links at the lemma level and due to the construction of an inconsistent morphemic structure. Our results indicate that letter transpositions at the morpheme boundary slows processing and changes the balance of the relative roles played by the whole word and by the embedded morphemes. When processing is slowed, the normally nondominant representations are able to play a relatively stronger role, such that the embedded morphemes play more of a role for the pseudocompounds, leading to a processing disadvantage relative to the controls, whereas the whole word representation plays more of a role for the compounds, leading to an attenuated processing advantage relative to the controls.

In sum, the current data shed light into the nature of processing compounds and words with embedded morphemes. So long as there are the two different levels (orthographic and morphological) operating, the effects observed in both the response time and accuracy analyses can all be generated by the differing directions of the two effects for the different word types. Moreover, the general assumption that there are two aspects to processing (i.e., orthographic and morphological) is consistent with other frameworks in the literature.

Given that our explanation suggests that compounds have semantic coherence between the constituents and the whole word whereas pseudocompounds lack this coherence, one would predict that semantic transparency should also influence the processing of the genuine compounds. Therefore, we examined the influence of semantic transparency on the correctly spelled compounds in Experiments 1 and 3. Semantic transparency ratings from human participants was obtained from the Large Database of English Compounds (Gagné, Spalding, & Schmidtke, in press). As in the other models, bigram frequency was entered as a covariate. The results support the prediction: Higher transparency was associated with greater ease of processing. In Experiment 1, semantic transparency was a successful predictor of both response time, z = −2.82, p = .005, and accuracy, z = 3.42, p = .001. Responses to compounds with higher transparency ratings were faster and more accurate than were responses to compounds with lower transparency ratings. In Experiment 3, semantic transparency was a successful predictor in the accuracy analysis, 2.30, p = .02, and was marginal in the response time analysis, z = −1.74, p = .08.

In general, our results are similar to other recent work investigating compound and pseudocompound word processing in showing that both kinds of words seem to trigger some form of morphological processing, but that the outcomes of that morphological processing attempt are different. For example, Gagné et al. (2018) found that using the compound or pseudocompound word as a prime for a constituent or pseudoconstituent led to highly robust priming effects. However, the priming effects were in opposite directions. In particular, the compound word prime led to facilitation of the constituent, but the pseudocompound word prime led to inhibition of the pseudoconstituent; teashop speeded the processing of tea, whereas carpet slowed the processing of car. In a study of typing latencies (Gagné & Spalding, 2016), compound words and pseudocompound words both led to differences from control words, showing effects of the morphemic or pseudomorphemic structure. Again, however, the compound and pseudocompound words differed from their matched controls in different ways. For example, compound words showed a large elevation in typing time exactly at the morpheme boundary, whereas pseudocompound words showed a smaller elevation in typing time, but the elevation began one letter before the pseudomorpheme boundary and extended into the second pseudoconstituent, likely because of competition among the activated (potential) morpheme representations and the true representation of the word.

Finally, in terms of the task itself, it is clear that the combination of the spelling error detection task with letter transpositions is a valuable experimental task for investigating morphological processing and, thus, might be a useful task to consider, especially when seeking converging evidence from other tasks such as lexical decision. The task leads to robust effects (e.g., of word type) with correctly spelled words and also leads to robust effects with incorrectly spelled words. This task has characteristics that make it a good task to compare with lexical decision, in that correct decisions require the participant to access the intended word, as in lexical decision. However, the spelling error detection task involves a decision process that has some advantages. For one, it is a much more natural task, especially for student populations, but probably for all literate populations: Outside of the laboratory, we people are quite commonly required to decide whether something is correctly spelled or not but are quite rarely presented with letter strings and asked whether or not the string is a word. In addition, comparisons of the different decisions (i.e., correctly vs. incorrectly spelled) in this task are more comparable than in lexical decision tasks (i.e., word vs. nonword), due to the fact that the nonword decision is, by hypothesis, quite open-ended: The participant has to somehow determine that the letter string does not match any word in the lexicon. Our findings suggest that the word access process and the decision process is largely preserved across the correctly and incorrectly spelled words.

Our results suggest that some attempt at morphological processing is obligatory, but this processing is only helpful when the true morphological structure of the word matches the apparent morphological structure, as in compounds. When there is a mismatch between the apparent and true morphological structures, as in pseudocompound words, processing is not helped. Thus, our findings reflect two effects. First, an orthographic effect that is facilitatory. Second, a morphemic effect that is facilitatory for compounds but inhibitory for pseudocompounds.

1  We thank an anonymous reviewer for this suggestion.

2  We thank Davide Crepaldi for raising this possibility.

3  We thank Davide Crepaldi for this suggestion.

Amenta, S., Marelli, M., & Crepaldi, D. (2015). The fruitless effort of growing a fruitless tree: Early morpho-orthographic and morpho-semantic effects in sentence reading. Journal of Experimental Psychology: Learning, Memory, and Cognition, 41, 1587–1596. 10.1037/xlm0000104

Andrews, S., & Davis, C. (1999). Interactive activation accounts of morphological decomposition: Finding the trap in mousetrap?Brain and Language, 68, 355–361. 10.1006/brln.1999.2067

Arcara, G., Semenza, C., & Bambini, V. (2014). Word structure and decomposition effects in reading. Cognitive Neuropsychology, 31, 184–218. 10.1080/02643294.2014.903915

Barr, D. J., Levy, R., Scheepers, C., & Tily, H. J. (2013). Random effects structure for confirmatory hypothesis testing: Keep it maximal. Journal of Memory and Language, 68, 255–278. 10.1016/j.jml.2012.11.001

Beyersmann, E., Castles, A., & Coltheart, M. (2011). Early morphological decomposition during visual word recognition: Evidence from masked transposed-letter priming. Psychonomic Bulletin & Review, 18, 937–942. 10.3758/s13423-011-0120-y

Beyersmann, E., McCormick, S. F., & Rastle, K. (2013). Letter transpositions within morphemes and across morpheme boundaries. The Quarterly Journal of Experimental Psychology, 66, 2389–2410. 10.1080/17470218.2013.782326

Beyersmann, E., Ziegler, J. C., Castles, A., Coltheart, M., Kezilas, Y., & Grainger, J. (2016). Morpho-orthographic segmentation without semantics. Psychonomic Bulletin & Review, 23, 533–539. 10.3758/s13423-015-0927-z

Brysbaert, M., & New, B. (2009). Moving beyond Kučera and Francis: A critical evaluation of current word frequency norms and the introduction of a new and improved word frequency measure for American English. Behavior Research Methods, Instruments & Computers, 41, 977–990. 10.3758/BRM.41.4.977

Chambers, S. M. (1979). Letter and order information in lexical access. Journal of Verbal Learning & Verbal Behavior, 18, 225–241. 10.1016/S0022-5371(79)90136-1

Christianson, K., Johnson, R. L., & Rayner, K. (2005). Letter transpositions within and across morphemes. Journal of Experimental Psychology: Learning, Memory, and Cognition, 31, 1327–1339. 10.1037/0278-7393.31.6.1327

Crepaldi, D., Rastle, K., Coltheart, M., & Nickels, L. (2010). Fell primes fall, but does bell prime ball? Masked priming with irregularly-inflected primes. Journal of Memory and Language, 63, 83–99. 10.1016/j.jml.2010.03.002

Crepaldi, D., Rastle, K., Davis, C. J., & Lupker, S. J. (2013). Seeing stems everywhere: Position-independent identification of stem morphemes. Journal of Experimental Psychology: Human Perception and Performance, 39, 510–525. 10.1037/a0029713

Diependaele, K., Sandra, D., & Grainger, J. (2005). Masked cross-modal morphological priming: Unravelling morpho-orthographic and morpho-semantic influences in early word recognition. Language and Cognitive Processes, 20, 75–114. 10.1080/01690960444000197

Duñabeitia, J. A., Perea, M., & Carreiras, M. (2007). Do transposed-letter similarity effects occur at a morpheme level? Evidence for morpho-orthographic decomposition. Cognition, 105, 691–703. 10.1016/j.cognition.2006.12.001

Fiorentino, R., & Fund-Reznicek, E. (2009). Masked morphological priming of compound constituents. The Mental Lexicon, 4, 159–193. 10.1075/ml.4.2.01fio

Fiorentino, R., & Poeppel, D. (2007). Compound words and structure in the lexicon. Language and Cognitive Processes, 22, 953–1000. 10.1080/01690960701190215

Gagné, C. L., & Spalding, T. L. (2004). Effect of relation availability on the interpretation and access of familiar noun-noun compounds. Brain and Language, 90, 478–486. 10.1016/S0093-934X(03)00459-0

Gagné, C. L., & Spalding, T. L. (2009). Constituent integration during the processing of compound words: Does it involve the use of relational structures?Journal of Memory and Language, 60, 20–35. 10.1016/j.jml.2008.07.003

Gagné, C. L., & Spalding, T. L. (2014a). Conceptual composition: The role of relational competition in the comprehension of modifier-noun phrases and noun-noun compounds. In B. H.Ross (Ed.), The psychology of learning and motivation (pp. 97–130). New York, NY: Elsevier.

Gagné, C. L., & Spalding, T. L. (2014b). Relational diversity and ease of processing for opaque and transparent English compounds. In F.Rainer, F.Gardani, H. C.Luschutzky, & W. U.Dressler (Eds.), Morphology and meaning (pp. 153–162). Amsterdam, the Netherlands: John Benjamins. 10.1075/cilt.327.10gag

Gagné, C. L., & Spalding, T. L. (2014c). Typing time as an index of morphological and semantic effects during English compound processing. Lingue E Linguaggio, XIII, 241–262.

Gagné, C. L., & Spalding, T. L. (2016). Effects of morphology and semantic transparency on typing latencies in English compound and pseudocompound words. Journal of Experimental Psychology: Learning, Memory, and Cognition, 42, 1489–1495. 10.1037/xlm0000258

Gagné, C. L., Spalding, T. L., & Nisbet, K. A. (2016). Processing English compounds: Investigating transparency. SKASE Journal of Theoretical Linguistics, 13, 2–22.

Gagné, C. L., Spalding, T. L., Nisbet, K. A., & Armstrong, C. (2018). Pseudo-morphemic structure inhibits, but morphemic structure facilitate, processing of a repeated free morpheme. Language, Cognition and Neuroscience, 33, 1252–1274. 10.1080/23273798.2018.1470250

Gagné, C. L., Spalding, T. L., & Schmidtke, D. (in press). LADEC: Large database of English compounds. Behaviour Research Methods.

Giraudo, H., & Grainger, J. (2001). Priming complex words: Evidence for supralexical representation of morphology. Psychonomic Bulletin & Review, 8, 127–131. 10.3758/BF03196148

Giraudo, H., & Voga, M. (2016). Words matter more than morphemes: An investigation of masked priming effects with complex words and non-words. Italian Journal of Linguistics, 28, 49–77.

Gomez, P., Ratcliff, R., & Perea, M. (2008). The overlap model: A model of letter position coding. Psychological Review, 115, 577–600. 10.1037/a0012667

Grainger, J., & Beyersmann, E. (2017). Edge-aligned embedded word activation initiates morpho-orthographic segmentation. In B.Ross (Ed.), Psychology of learning and motivation (pp. 285–317). Cambridge, MA: Academic Press.

Ji, H., Gagné, C. L., & Spalding, T. L. (2011). Benefits and costs of lexical decomposition and semantic integration during the processing of transparent and opaque English compounds. Journal of Memory and Language, 65, 406–430. 10.1016/j.jml.2011.07.003

Jones, M. N., & Mewhort, D. J. (2004). Case-sensitive letter and bigram frequency counts from large-scale English corpora. Behavior Research Methods, Instruments, & Computers, 36, 388–396. 10.3758/BF03195586

Kazanina, N., Dukov-Zheleva, G., Geber, D., Kharlamov, V., & Tonciulescu, K. (2008). Decomposition into multiple morphemes during lexical access: A masked priming study of Russian nouns. Language and Cognitive Processes, 23, 800–823. 10.1080/01690960701799635

Levelt, W. J., Roelofs, A., & Meyer, A. S. (1999). A theory of lexical access in speech production. Behavioral and Brain Sciences, 22, 1–38. 10.1017/S0140525X99001776

Libben, G. (1994). How is morphological decomposition achieved?Language and Cognitive Processes, 9, 369–391. 10.1080/01690969408402124

Libben, G., Curtiss, K., & Weber, S. (2014). Psychocentricity and participant profiles: Implications for lexical processing among multilinguals. Frontiers in Psychology, 5, 557. 10.3389/fpsyg.2014.00557

Libben, G., Gibson, M., Yoon, Y. B., & Sandra, D. (2003). Compound fracture: The role of semantic transparency and morphological headedness. Brain and Language, 84, 50–64. 10.1016/S0093-934X(02)00520-5

Longtin, C. M., Segui, J., & Hallé, P. A. (2003). Morphological priming without morphological relationship. Language and Cognitive Processes, 18, 313–334. 10.1080/01690960244000036

MacKay, D. G. (1972). Input testing in the detection of misspellings. The American Journal of Psychology, 85, 121–127. 10.2307/1420967

MacKay, D. G. (1992). Awareness and error detection: New theories and research paradigms. Consciousness and Cognition, 1, 199–225. 10.1016/1053-8100(92)90061-E

Marelli, M., Amenta, S., Morone, E. A., & Crepaldi, D. (2013). Meaning is in the beholder’s eye: Morpho-semantic effects in masked priming. Psychonomic Bulletin & Review, 20, 534–541. 10.3758/s13423-012-0363-2

Matuschek, H., Kliegl, R., Vasishth, S., Baayen, H., & Bates, D. (2017). Balancing Type I error and power in linear mixed models. Journal of Memory and Language, 94, 305–315. 10.1016/j.jml.2017.01.001

Norris, D., Kinoshita, S., & van Casteren, M. (2010). A stimulus sampling theory of letter identity and order. Journal of Memory and Language, 62, 254–271. 10.1016/j.jml.2009.11.002

O’Connor, R. E., & Forster, K. I. (1981). Criterion bias and search sequence bias in word recognition. Memory & Cognition, 9, 78–92. 10.3758/BF03196953

Perea, M., & Carreiras, M. (2006). Do transposed-letter effects occur across lexeme boundaries?Psychonomic Bulletin & Review, 13, 418–422. 10.3758/BF03193863

Perea, M., & Lupker, S. J. (2004). Can CANISO activate CASINO? Transposed-letter similarity effects with non-adjacent letter positions. Journal of Memory and Language, 51, 231–246. 10.1016/j.jml.2004.05.005

Rabe-Hesketh, S., & Skrondal, A. (2012). Multilevel and longitudinal modeling using Stata. College Station, TX: Stata Press Publication.

Rastle, K., Davis, M. H., & New, B. (2004). The broth in my brother’s brothel: Morpho-orthographic segmentation in visual word recognition. Psychonomic Bulletin & Review, 11, 1090–1098. 10.3758/BF03196742

Rueckl, J. G., & Rimzhim, A. (2011). On the interaction of letter transpositions and morphemic boundaries. Language and Cognitive Processes, 26, 482–508. 10.1080/01690965.2010.500020

Sánchez-Gutiérrez, C., & Rastle, K. (2013). Letter transpositions within and across morphemic boundaries: Is there a cross-language difference?Psychonomic Bulletin & Review, 20, 988–996. 10.3758/s13423-013-0425-0

Sandra, D. (1994). The morphology of the mental lexicon: Internal word structure viewed from a psycholinguistic perspective. Language and Cognitive Processes, 9, 227–269. 10.1080/01690969408402119

Spalding, T. L., & Gagné, C. L. (2011). Relation priming in established compounds: Facilitation?Memory & Cognition, 39, 1472–1486. 10.3758/s13421-011-0112-1

Spalding, T. L., & Gagné, C. L. (2014). Relational diversity affects ease of processing even for opaque English compounds. The Mental Lexicon, 9, 48–66. 10.1075/ml.9.1.03spa

Taft, M. (2003). Morphological representation as a correlation between form and meaning. In E. M. H.Assink & D.Sandra (Eds.), Reading complex words: Cross-language studies (pp. 113–137). Amsterdam, The Netherlands: Kluwer Academic. 10.1007/978-1-4757-3720-2_6

Taft, M., & Forster, K. I. (1975). Lexical storage and retrieval of prefixed words. Journal of Verbal Learning & Verbal Behavior, 14, 638–647. 10.1016/S0022-5371(75)80051-X

Taft, M., & Forster, K. (1976). Lexical storage and retrieval of polymorphemic and polysyllabic words. Journal of Verbal Learning & Verbal Behavior, 15, 607–620. 10.1016/0022-5371(76)90054-2

Taft, M., & Nillsen, C. (2013). Morphological decomposition and the transposed-letter (TL) position effect. Language and Cognitive Processes, 28, 917–938. 10.1080/01690965.2012.679662

Submitted: June 18, 2018 Revised: June 11, 2019 Accepted: June 12, 2019