Morphology is crucial to word formation. For example, “plays,” “player,” and “playground” can be constructed from “play.” The role of morphemes in complex word recognition has been studied extensively (Amenta & Crepaldi, 2012). In the last decade, considerable evidence has been accumulated to support the proposal that morphemic form alone is sufficient to trigger morphological decomposition (Rastle & Davis, 2008). Specifically, in a masked priming experiment, stem word recognition was facilitated not just by transparent primes (morphemes related to word meanings; e.g., cleaner–CLEAN), but also by opaque/pseudocomplex primes (morphemes unrelated to word meanings; e.g., corner–CORN), whereas letter-sharing had no effects (e.g., brothel–BROTH, since “-el” is not an English suffix; Rastle, Davis, & New, 2004). Opaque word priming has also been observed in Chinese (Tsang & Chen, 2013a) and French (Longtin, Segui, & Hallé, 2003), supporting morpho-orthographic decomposition across different languages.

Although it is widely accepted that morphemic forms are activated to trigger rapid morphological decomposition, it remains equivocal whether morpho-semantics is available early. Longtin et al. (2003) and Rastle et al. (2004) argued against early morpho-semantic activation, on the basis of their findings that transparent and opaque primes facilitated word recognition equally. In contrast, Diependaele, Sandra, and Grainger (2005, 2009) and Feldman, O’Connor, and Moscoso del Prado Martín (2009) observed stronger masked priming by transparent than by opaque words, which was taken as evidence for fast morpho-semantic activation. Other researchers have studied the issue with ambiguous morphemes and discovered stronger masked priming when the primes and targets shared morphemic forms and meanings (e.g., invalid–incorrect), as compared to form-sharing only (e.g., inside–incorrect; Zhou, Marslen-Wilson, Taft, & Shu, 1999). Moreover, the early stage of morphemic ambiguity resolution was sensitive to various morpho-semantic variables, such as meaning frequency and contextual constraints (Taft & Nguyen-Hoan, 2010; Tsang & Chen, 2013a, 2013b).

To accommodate early morpho-semantic activation, some models incorporate explicit morpho-semantic representations. For instance, Taft and Nguyen-Hoan (2010) proposed that morphemes are represented not just at the form level, but also at the lemma level, an abstract layer of representations coding consistent form–meaning linkages. Thus, morpho-semantic priming can be attributed to repeated activation of the same lemmas. Crepaldi, Rastle, Coltheart, and Nickels (2010) also adopted the lemma idea, although they restricted lemma sharing only to inflected words (e.g., “fall–fell,” “cat–cats”), but not to derived words (and by extension, compound words), because the latter cases often involve subtle changes in meanings and word classes (e.g., “critic–critical,” “baseball–baseline”). In the hybrid dual-route model, Diependaele et al. (2009) proposed that morpho-orthographic decomposition and holistic processing operate simultaneously. Morphemic meanings can be activated postlexically and rapidly via the morphological route and the lexical route, which send excitatory feedback to influence morphological processing, accounting for the morpho-semantic effects in masked priming experiments.

Although these models assume different mechanisms of morphological processing, they are consistent in suggesting the possibility that morphemic meanings can be activated automatically and rapidly, not just for transparent, but also for opaque words. For instance, the meanings of “corn” and “clean” can be activated once “corner” and “cleaner” are decomposed on the basis of surface morpho-orthography. Although for opaque words the activated morphemic meanings must eventually be inhibited by whole-word meanings to avoid miscomprehension, the inhibition might have been ineffective in previous masked priming experiments for several reasons. First, stem words were used as targets, which completely matched with the decomposed morphemes, and typically had higher frequency than the opaque primes. This strengthened the facilitation from morphemic meanings and, comparatively, the inhibition from whole words might have been too weak, especially given a short stimulus onset asynchrony (<50 ms). Thus, it is possible that within the temporal window reflected by masked priming, morphemic meanings could be highly active for both transparent and opaque words, which means that the absence of a semantic transparency effect (Longtin et al., 2003; Rastle et al., 2004) could not be taken as indisputable evidence against early morpho-semantic activation. In this study, we reexamined early morpho-semantic activation during opaque word processing in two masked priming experiments.

In Experiment 1, we compared the masked priming effects of opaque primes on transparent targets against those of transparent primes on opaque targets. If only morphemic form was activated, the strength of facilitation should be identical across the two conditions, because there were equal amounts of morpho-orthographic sharing. In contrast, if early morpho-semantic activation occurred, facilitation of the opaque targets should diminish or disappear, as compared to the transparent targets, because the morphemic meanings activated by the constituent morphemes in prime words would be related only to the transparent targets, not to the opaque targets. This should strengthen the inhibition from whole words in the opaque target condition. In Experiment 2, we used target words that were unrelated to the whole opaque primes, but were semantically related to the initial constituent morphemes. Any facilitation observed in this condition would provide direct evidence for morpho-semantic activation even in opaque words.

Experiment 1

Method

Participants

A group of 48 undergraduates from the Chinese University of Hong Kong were recruited. They were native Cantonese speakers and had normal or corrected-to-normal vision.

Materials

Twenty-four Chinese morphemes were used to construct pairs of transparent and opaque bimorphemic words (critical morphemes were always at the initial position). For transparent compound words, the critical morphemes contributed to the word meanings (e.g., 雷雨 thunderstorm, meaning “thunder–rain” literally). The opaque words were either loan words from other languages (e.g., 雷達 radar, meaning “thunder–arrive” literally) or words used only in highly specific contexts (e.g., 點心 Dim Sum, meaning “dot–heart” literally). Twelve loan words bore phonological similarity to their source English words. Although this might increase their phonological salience, it could not explain any morpho-semantic effects observed, because phonological sharing, like moprho-orthographic sharing, was identical across conditions. Twenty-four bimorphemic control words were also created with morphemes not used in the transparent and opaque conditions. These control words were fully transparent (e.g., 飯盒 lunch box, meaning “rice–box” literally). The words in different conditions were matched for log-transformed word frequency (Cai & Brysbaert, 2010), frequency of the second morphemes (Research Centre for Humanities Computing, 2003), and number of strokes (Table 1).

Table 1 Sample materials and their properties
Full size table

Four conditions were formed by crossing target type (transparent vs. opaque) and prime type (form-related vs. unrelated) as independent variables (when transparent words were targets, opaque words served as the form-related primes; similarly, for opaque targets, transparent words served as the form-related primes). This design prevented any frequency difference between the primes and targets. The unrelated primes were Chinese bimorphemic words that were unrelated to the targets at both morphological and lexical levels. To avoid item repetition, four lists were constructed such that each target appeared only once within a list and all possible prime–target pairings were exhausted across lists. Six items were presented in a list per condition. For the lexical decision task, 24 nonword targets were created by combining Chinese morphemes in a noninterpretable way (e.g., 河亮, meaning “river–bright” literally). Thus, meaning retrieval was necessary to perform the lexical decision. All of the nonwords were preceded by real-word primes, half of which shared morphemes with the targets, such that morpheme-sharing was not informative as to the lexical status of targets. The same set of nonwords was added to each experimental list (i.e., 48 items in each list).

Procedure

Five to eight participants performed in each experimental session. They were arranged in computer booths in order to present materials individually and to minimize interference from other participants. Each trial began with a fixation cross in the center of the screen for 500 ms, which was replaced by a random stroke pattern as the forward mask. The prime word (in PMingLiU font) was then presented for 40 ms. The target (in DFKai-SB font) appeared immediately after the prime and stayed on the screen until participants responded, or for 2,000 ms when no responses were made. Participants were instructed to decide as quickly and as accurately as possible whether the target was a real Chinese word by pressing corresponding keys on the keyboard. They were not informed about the presence of the primes, but were told that they might see a “flash of symbols” before the targets. The combination of a forward mask and a short presentation duration ensured that participants were unaware of the primes. Twelve practice trials were provided. The participants were randomly assigned to the four lists (12 participants in each list). Items within a list were presented randomly, and the experiment lasted about 15 min.

Results and discussion

Trials with incorrect responses (9.2 %) or extreme reaction times (±2.5 SDs; 1.3 %) were discarded from further analyses. Table 2 displays the mean reaction times and error rates in each condition for the remaining data. A linear mixed-effect model (lme; Baayen, Davidson, & Bates, 2008) and a mixed logit model (Jaeger, 2008) were used to analyze the reaction time data and error rates, respectively. Target type (transparent vs. opaque) and prime type (form related vs. unrelated) were treated as fixed effects, and subject and item were treated as random effects. A treatment coding was adopted such that opaque targets were compared with transparent targets, and related primes were compared with unrelated primes. Following Barr, Levy, Scheepers, and Tily (2013), random intercepts, random slopes, and their correlations were included in the random effect structure. For the lme, given a relatively large data set (over 1,000 valid observations), the t distribution should approximate a normal distribution. Thus, any effects with t > 2 were considered to be significant at the .05 level. For the mixed logit model, we report the Wald Z values and p values obtained by Laplace approximation.

Table 2 Mean reaction times (standard errors of the means in parentheses) and error rates in Experiment 1
Full size table

For error rates, only the main effect of target type was significant (Wald z = 3.11, p < .005). Fewer errors were committed for transparent targets. The prime type effect and the interaction were nonsignificant (Wald zs = 1.22, n.s.). For reaction times, responses were faster for transparent than for opaque targets (b = 95.36, SE = 29.87; t = 3.19, p < .01). The main effect of prime relatedness was nonsignificant (b = 6.88, SE = 23.72; t < 1). However, the main effects were characterized by a significant interaction between target and prime type (b = 56.55, SE = 23.83; t = 2.37, p < .05). To further specify the nature of this interaction, a separate lme test was conducted for each target type. A significant prime type effect was obtained for transparent targets (b = 49.23, SE = 14.75; t = 3.34, p < .01), but not for opaque targets (b = 8.65, SE = 23.90; t < 1).

Although opaque primes facilitated transparent target recognition, which was consistent with previous findings with stem word targets (Longtin et al., 2003; Rastle et al., 2004), transparent primes failed to influence the processing of opaque targets.Footnote 1 Similar results were recently obtained in an event-related potential experiment by Morris, Grainger, and Holcomb (2013), in which complex nonwords facilitated the recognition of transparent targets (e.g., “farmity–farmer”), but had no effects for opaque targets (e.g., “cornity–corner”). This asymmetry can be readily explained if we assume fast morpho-semantic activation once decomposition based on surface morphemic form is complete. For opaque primes, the morphemic meanings activated were related to the meanings of the whole transparent targets, producing strong facilitation. However, the morphemic meanings of the transparent primes were irrelevant to the opaque targets. Instead, they might be more strongly connected to other transparent words containing the same morphemes, because many transparent words were more frequent than the opaque ones, creating competition with opaque target activation that might cancel out facilitation due to the sharing of morphemic forms. This explains why morphological priming was nonsignificant for opaque targets, even though morpho-orthographic sharing should still be at work.

The idea of fast morpho-semantic activation also implies that the constituent morphemes should influence the recognition of semantically related targets at the early stage of processing reflected by masked priming (e.g., facilitation of “bread” by “butterfly”). This was tested in Experiment 2.

Experiment 2

Method

Participants

A group of 30 participants from the same population as in Experiment 1 were recruited. None of them had participated in other experiments or pilot testing.

Materials

The transparent, opaque, and unrelated words in Experiment 1 were used as primes in Experiment 2. Targets were transparent bimorphemic Chinese words that were semantically related to the initial morphemes, but not to the whole opaque words. Therefore, facilitation by opaque primes could only be attributed to morpho-semantic activation, but not to lexical semantics. Given that morphemes contribute to the meanings of transparent words, the targets were also related to the transparent primes. Semantic relatedness was rated by 20–27 pilot participants on a 6-point Likert scale (1 = strongly unrelated, 6 = strongly related). For example, the target word 閃光 flash was related to the morpheme 雷 thunder (mean rating = 4.78), and also to the transparent prime 雷雨 thunderstorm (mean = 4.41). But it was unrelated to both the opaque (雷達 radar, mean = 2.47) and the unrelated (飯盒 lunchbox, mean = 2.42) primes. Three experimental lists were created to prevent target repetition. All prime–target combinations were exhausted across lists. Eight items were presented per condition, and eight fillers with unrelated primes and targets were added to balance the numbers of related and unrelated trials. Thirty-two nonword items similar to those in Experiment 1 were also added for lexical decisions.

Procedure

The procedure was identical to that of Experiment 1, with ten participants in each list.

Results and discussion

Trials with incorrect responses (5.1 %) or extreme reaction times (±2.5 SD; 1.5 %) were discarded from further analyses. Table 3 displays the mean reaction times and error rates in each condition. The data were analyzed as in Experiment 1. Prime type (transparent vs. opaque vs. control) was treated as fixed effect, and subject and item were treated as random effects while maintaining a maximal random-effect structure.

Table 3 Mean reaction times (standard errors of the means in parentheses) and error rates in Experiment 2
Full size table

In the first analysis, a simple-effect model was used to compare performance in the two morpheme-sharing conditions against the unrelated control. Reaction times were significantly faster in both the transparent (b = 57.02, SE = 19.55; t = 2.92, p < .05) and opaque (b = 39.65, SE = 15.92; t = 2.49, p < .05) conditions. In the second analysis, a linear contrast (transparent vs. opaque, opaque vs. control) was set up to examine the pairwise differences across conditions. Reaction times were comparable in the transparent and opaque conditions (b = 17.38, SE = 16.76; t = 1.04), whereas a significant difference emerged between the opaque and the control conditions (b = 39.65, SE = 15.92; t = 2.49, p < .05). We found no significant differences in error rates (Wald z = 0.36 to 1.60).

Masked morphological priming was produced by the transparent and opaque primes, with nonsignificant difference between the two conditions. Most importantly, the effects in the opaque prime condition could only be attributed to morpho-semantic activation, because the primes were unrelated to the targets at the lexical level. These results provided direct evidence for early morpho-semantic activation for opaque words.

General discussion

Previous masked priming experiments that compared transparent and opaque words (Diependaele et al., 2005; Longtin et al., 2003; Rastle et al., 2004) had typically considered nonsignificant semantic transparency effects as evidence against early morpho-semantic activation. However, as we discussed in the introduction, it is possible that opaque words could temporarily activate morphemic meanings as strongly as transparent words did, because of the masked priming procedure and the choice of materials. In other words, both morphemic form and meanings are activated at the early stages of processing. In light of this possibility, we reexamined masked morpho-semantic priming by opaque words in two experiments. Experiment 1 showed that masked morphological priming cannot be explained solely by the sharing of morphemic forms, because the priming effect was nonsignificant in the transparent prime–opaque target condition (despite a typical opaque prime–transparent target effect being found). In Experiment 2, opaque primes could facilitate the recognition of targets that were related to morphemic meanings but unrelated to whole words, providing direct evidence of early morpho-semantic activation for opaque words.

The issue of early morpho-semantic activation is not trivial, because it provides insights to how quickly language users can retrieve meanings from linguistic inputs (Feldman et al., 2009). In light of these results and our studies on morphemic ambiguity (Tsang & Chen, 2013a, 2013b), we suggest that morphemic meanings can be activated very quickly, even before the involvement of conscious awareness. For transparent words, the morphemic meanings activated contribute to the meanings of the whole words, and morphological integration would not interfere with lexical access. In contrast, for opaque words, although integration is attempted, the outputs are noninterpretable, causing troubles in word recognition (Ji, Gagné, & Spalding, 2011). Lavric, Rastle, and Clapp (2011) further suggested that failures of morphological integration would trigger a semantic reanalysis based on whole words, as evidenced by a stronger N400 for opaque than for transparent words. This reanalysis is costly and time-consuming, cancelling out the morpho-orthographic facilitation at later stages of processing that is reflected by cross-modal priming (Marslen-Wilson, Tyler, Waksler, & Older, 1994). It also explains why the morphemes in opaque words could facilitate the recognition of semantic targets in masked priming (Exp. 2), but not in unmasked priming (Sandra, 1990). On the other hand, the limits of morpho-semantic activation require further investigation. In particular, it remains to be tested whether morpho-semantic effects would persist when the morphemes were presented in isolation (i.e., as individual words), given that masked priming is usually weaker when the primes and targets are related in lexical semantics (Rastle, Davis, Marslen-Wilson, & Tyler, 2000).

The morpho-semantic effects observed in this study are consistent with models that assume separate morpho-semantic representations. Judging which model best fits the empirical data will not be easy, but our findings can provide important constraints for model selection. First, morphemic meanings are activated even for opaque words, suggesting automatic morpho-semantic activation after morpho-orthographic decomposition. Second, quick morpho-semantic activation has been observed for both compound (this study) and derived (Diependaele et al., 2005) words. Therefore, the model by Crepaldi et al. (2010) might need to be extended to accommodate morpho-semantic relationships other than inflections. On the other hand, the lemma model of Taft and Nguyen-Hoan (2010) and the dual-route model (Diependaele et al., 2009) are both general enough to account for the present findings. However, the dual-route model may need further justifications to explain why morphological effects emerged earlier than lexical ones (Tsang & Chen, 2010; Zhou et al., 1999), despite the assumption of a parallel lexical route for word recognition. Further experiments will be needed to verify these models.

Future research will also be needed to investigate why early semantic transparency effects have been found in some studies (Diependaele et al., 2005, 2009; Feldman et al., 2009) but not in others (Longtin et al., 2003; Rastle et al., 2004). As we briefly mentioned in the introduction, this finding may be related to the frequency of constituent morphemes. For high-frequency morphemes, the strongly activated morphemic meanings in opaque words can temporarily resist inhibition from whole words. The strengths of morpho-semantic activation would then be comparable in transparent and opaque conditions, resulting in nonsignificant differences between the two conditions. In contrast, for low-frequency morphemes, strong competitions between morphemic meanings and whole words would occur for opaque words, but not for transparent words. In this case, transparent primes would lead to stronger facilitation than opaque primes, creating the semantic transparency effect observed.

Finally, it will be important to replicate the present findings in other languages, since written Chinese has unique properties that may have strengthened the morpho-semantic effects. For instance, the linkage between form and meaning may be stronger in Chinese than in languages with alphabetic scripts (Zhou et al., 1999). The physical spaces separating individual morphemes may also simplify morpho-orthographic decomposition and accelerate morpho-semantic processing. Yet, early morpho-semantic activation has also been demonstrated in other languages (Diependaele et al., 2005; Feldman et al., 2009; Taft & Nguyen-Hoan, 2010), which leads us to believe that morpho-semantic activation during opaque word processing is language-universal.