The locus of morphemic representation in L1 processing of complex words

Although it is now widely accepted that morphology plays a significant role in L1 visual recognition of complex words, there is no consensus among researchers regarding the locus of morphemic representation. Three competing models have been proposed. The sublexical model (Rastle, Davis & New, Reference Rastle, Davis and New2004; Taft & Forster, Reference Taft and Forster1976; Taft, Reference Taft1994) posits that a complex word is automatically decomposed into its morphemic constituents purely based on orthography prior to the activation of its whole-word lexical access, also known as the ‘morpho-orthographic’ decomposition account. In this model, morphemic representations are contacted before whole-word representations, and morphological decomposition is used to facilitate lexical access of the whole word. The supralexical model (Giraudo & Grainger, Reference Giraudo and Grainger2000; Giraudo & Grainger, Reference Giraudo and Grainger2001), on the other hand, posits that morphemic representations are contacted subsequent to the recognition of the whole word, and morphological processing is constrained by the ‘morpho-semantic’ properties of the complex word. The hybrid model (Diependaele et al., Reference Diependaele, Sandra and Grainger2005; Diependaele, Sandra & Grainger, Reference Diependaele, Sandra and Grainger2009) proposes a double locus of morphological representation as opposed to a single one, and includes both a sublexical morpho-orthographic processing system and a supralexical morpho-semantic processing system, operating in parallel, in early visual recognition.

Empirical evidence from masked priming has been used to argue for or against each of these models. Giraudo and Grainger (Reference Giraudo and Grainger2000) found that the magnitude of morphological priming relative to orthographic controls was modulated by the surface frequency but not the cumulative root frequency (i.e., the summed frequency of all derived words sharing the same root) of the prime words. The absence of cumulative root frequency effects was used to argue against the sublexical hypothesis, and the presence of surface frequency effects was used to provide support for the supralexical model. Giraudo and Grainger (Reference Giraudo and Grainger2001), in Experiment 1, found that derived suffixed words and their roots were equally effective primes for other suffixed derivations of the same roots in French, irrespective of prime exposure durations (43 ms and 57 ms). This finding runs counter to the sublexical hypothesis based on the assumption that the prelexical parsing of derived word primes should require extra computation and thus more time than the processing of free roots. In Experiment 2, they used derived suffixed words as targets preceded by derived suffixed word primes sharing the same roots (e.g., laitage [milk product] – laitier [milkman], which is like creation – creator) or monomorphemic word primes containing a pseudoroot (e.g., laitue [lettuce] – laitier [milkman], which is like costume – costly). They found that the former produced a priming effect, but the latter did not. Giraudo and Grainger used this finding as evidence against the blind morphological decomposition account.

Evidence in support of the sublexical account has been obtained from masked priming studies that manipulated prime-target relatedness across three conditions: (a) morphologically, semantically, and orthographically related (+M+S+O; e.g., cleaner-CLEAN); (b) morphologically and orthographically related, but semantically unrelated/opaque (+M-S+O; e.g., department-DEPART); and (c) orthographically related only (-M-S+O; e.g., brothel-BROTH). It has been consistently found in a number of studies that significant masked priming occurs in conditions (a) and (b), but not in condition (c). That is, facilitative masked priming occurs with both semantically transparent and opaque primes, while mere orthographic overlap is not sufficient to produce priming effects (Fiorentino & Fund-Reznicek, Reference Fiorentino and Fund-Reznicek2009, on compound words; Rastle & Davis, Reference Rastle and Davis2008, for a qualitative meta-analysis of 19 empirical studies on derived words). More importantly, the priming effects for the transparent and opaque primes are statistically equivalent in the majority of the studies (Rastle & Davis, Reference Rastle and Davis2008). This pattern of masked priming effects has been taken as evidence for an early stage of automatic and blind decomposition that is independent of semantic transparency but largely guided by what orthographically appears to be morphological structure, which is known as the morpho-orthographic segmentation hypothesis (Rastle et al., Reference Rastle, Davis and New2004).

A few recent studies, however, have observed a semantic transparency effect in early visual masked priming (Diependaele et al., Reference Diependaele, Sandra and Grainger2005; Diependaele et al., Reference Diependaele, Sandra and Grainger2009; Diependaele et al., Reference Diependaele, Duñabeitia, Morris and Keuleers2011; Feldman, O’Connor & Moscoso del Prado Martín, Reference Feldman, O’Connor and Moscoso del Prado Martín2009; Feldman, Kostić, Gvozdenović, O’Connor & Moscoso del Prado Martín, Reference Feldman, Kostić, Gvozdenović, O’Connor and Moscoso del Prado Martín2012; Morris, Frank, Grainger & Holcomb, Reference Morris, Frank, Grainger and Holcomb2007). Diependaele and colleagues (Reference Diependaele, Sandra and Grainger2005, Reference Diependaele, Sandra and Grainger2009, Reference Diependaele, Duñabeitia, Morris and Keuleers2011), for example, in their investigation of masked priming with derived words (including both prefixed and suffixed), found the same pattern of priming effects, i.e., significant facilitative priming with both transparent and opaque primes, but stronger priming for the transparent than for the opaque primes. In a statistical meta-analysis of the same studies included in Rastle and Davis (Reference Rastle and Davis2008), Feldman and colleagues (Reference Feldman, O’Connor and Moscoso del Prado Martín2009) showed a significant advantage of transparent primes over opaque primes, although this advantage is not always significant in individual studies. The pattern of graded priming, across semantically transparent and opaque primes, seems to be difficult to reconcile with the pure sublexical morpho-orthographic processing account. Diependaele and colleagues (Reference Diependaele, Sandra and Grainger2009) then proposed a hybrid account that includes both sublexical morpho-orthographic processing and supralexical morpho-semantic processing, operating in parallel in early visual recognition. The advantage of transparent primes over opaque ones can be explained by this account in that the constituent priming from transparent words benefits not only from sublexical morpho-orthographic activation, but also from top-down supralexical morpho-semantic facilitation, while the constituent priming from opaque words can only benefit from one source (i.e., sublexical morpho-orthographic activation that decays quickly).

Major issues in L2 processing of morphologically complex words

Research on L2 processing of complex words has been an extension of L1 morphological processing, and the major research question has been whether morphology also plays a role in L2 complex word processing. It is hotly debated whether L2 morphological processing differs from L1 processing, and, if so, how and why. Some researchers have argued that L2 and L1 morphological processing are fundamentally different and the differences cannot be sufficiently explained by L1 transfer or cognitive resource limitations (e.g., Clahsen et al., Reference Clahsen, Felser, Neubauer, Sato and Silva2010). Across a number of studies on the processing of inflected words, Clahsen and colleagues found significant masked stem priming effects in NSs but no such effect for L2 learners (English regular past-tense -ed with Arabic L1 speakers in Clahsen et al., Reference Clahsen, Balkhair, Schutter and Cunnings2013; Turkish regular verb inflection with various L1 speakers in Kirkici & Clahsen, Reference Kirkici and Clahsen2013; German inflected forms with Polish L1 speakers in Neubauer & Clahsen, Reference Neubauer and Clahsen2009; English regular past-tense -ed with Chinese, Japanese, and German L1 speakers in Silva & Clahsen, Reference Silva and Clahsen2008). They argue that “adult L2 learners are less sensitive to morphological structure than native speakers and rely more on lexical storage than on morphological parsing during processing” (Clahsen et al., Reference Clahsen, Felser, Neubauer, Sato and Silva2010, p. 21).

Some other researchers have argued, however, that L2 and L1 morphological processing share the same mechanisms, and the differences between the L1 and L2 processing performance can be attributed to L1 transfer (Basnight-Brown et al., Reference Basnight-Brown, Chen, Hua, Kostić and Feldman2007; Portin et al., Reference Portin, Lehtonen, Harrer, Wande, Niemi and Laine2008) or domain-general cognitive resource limitations (McDonald, Reference McDonald2006). There has been evidence from research involving both inflection and derivation that partially supports this position (Diependaele et al., Reference Diependaele, Duñabeitia, Morris and Keuleers2011; Feldman et al., Reference Feldman, Kostić, Basnight-Brown, Filipović-Đurđević and Pastizzo2010; Gor & Cook, Reference Gor and Cook2010; Portin et al., Reference Portin, Lehtonen and Laine2007). With regard to inflection, Basnight-Brown et al. (Reference Basnight-Brown, Chen, Hua, Kostić and Feldman2007) found that both Serbian–English and Chinese–English bilinguals process English regular verbs similarly and like NSs in a cross-modal priming study. In addition, Feldman and colleagues (Reference Feldman, Kostić, Basnight-Brown, Filipović-Đurđević and Pastizzo2010) found reliable masked priming effects for English regular past tense -ed in both NSs and NNSs (Serbian L1s). L2 evidence on derivations comes from Diependaele et al. (Reference Diependaele, Duñabeitia, Morris and Keuleers2011), who compared masked priming effects of English derived words among NSs and two groups of NNSs (Spanish and Dutch L1s). They found that both native and bilingual participants (regardless of L1s) showed similar patterns of priming effects, and therefore suggested that late bilinguals largely use the same processing strategies as NSs do.

The issue of whether the mechanism underlying L2 processing of complex words is similar to or distinct from L1 processing is far from settled. It is important to note that these competing views have been informed exclusively by research on inflections and derivations. In a recent study, Kirkici and Clahsen (Reference Kirkici and Clahsen2013) tried to draw attention to “the morphological differences between inflectional and derivational processes” (p. 776), and found psycholinguistic evidence for different priming patterns for inflection and derivation within the L2 group (i.e., significant priming for derived forms, but no priming for inflected forms, in L2 learners of Turkish). These different patterns of priming effects for inflection and derivation suggest that there may be different mechanisms underlying L2 processing of complex words depending on the type of morphology. The question whether the underlying mechanism for L2 processing of compounds is the same as or different from L2 processing of inflected and/or derived words warrants empirical investigation.

Empirical research on L1 and L2 processing of compound words

A comprehensive review of previous empirical research on L1 compound processing using different experimental paradigms is beyond the scope of this present study. Instead, our focus is on masked priming studies that have investigated the role of semantic transparency since it is pertinent to the discussion of the locus of morphemic representation. The two published studies on English compound processing that used the masked priming paradigm (Fiorentino & Fund-Reznicek, Reference Fiorentino and Fund-Reznicek2009; Shoolman & Andrews, Reference Shoolman, Andrews, Kinoshita and Lupker2003), did not find an effect of semantic transparency. Using compounds as targets, and their monomorphemic constituents as primes, Shoolman and Andrews (Reference Shoolman, Andrews, Kinoshita and Lupker2003) found that both the first and second constituents primed the compounds regardless of semantic transparency. Arguing that Shoolman and Andrews’ (Reference Shoolman, Andrews, Kinoshita and Lupker2003) overt presentation of compounds as targets leaves it unclear whether the observed priming effects reflect automatic processing, Fiorentino and Fund-Reznicek (Reference Fiorentino and Fund-Reznicek2009) used compounds as primes and their constituents as targets. They found significant and statistically equivalent facilitative priming effects in the transparent and opaque conditions but no priming in the orthographic overlap condition, in both the word-initial and the word-final overlap positions. They concluded that the morpho-orthographic segmentation independent of semantic transparency that has been consistently found in the processing of affixed words can be generalized to compound processing. In Chinese compound processing, however, semantic transparency has been found to play a role in visual masked priming. Peng, Liu, and Wang (Reference Peng, Liu, Wang, Wang, Inhoff and Chen1999), in a masked priming experiment with compounds varying in semantic transparency presented as primes for 56 ms, found significant priming effect only for semantically transparent compounds, not for opaque compounds. Liu and Peng (Reference Liu, Peng and Chen1997) found similar results in a visual priming task with an SOA of 86ms. These findings point to early decomposition for transparent compounds, but whole-word access for opaque compounds with possible supralexical decomposition in Chinese compound processing.

When it comes to L2 compound processing, there have been much fewer studies. The line of the studies that investigated cross-language activation in L2 compound processing (Cheng, Wang & Perfetti, Reference Cheng, Wang and Perfetti2011 with Chinese–English bilingual children; Ko, Wang & Kim, Reference Ko, Wang and Kim2011 with adult Korean–English bilinguals; Wang, Lin & Gao, Reference Wang, Lin and Gao2010 with adult Chinese–English bilinguals) used lexical decision tasks without priming and concluded that they found evidence for compound decomposition in L2 processing. Lemhöfer, Koester, and Schreuder (2011) found that both NSs and adult German–Dutch bilinguals responded faster to Dutch compounds that contained an orthotactic cue in a lexical decision task, which provides evidence for compound decomposition in L2 processing. De Cat, Klepousniotou, and Baayen (Reference De Cat, Klepousniotou and Baayen2015) examined L2 processing of English transparent, low-frequency noun-noun compounds in native Spanish and German speakers and found evidence for compound decomposition by advanced NNSs with L1 interference effects. Most relevant to the present study in terms of design, Ko (Reference Ko2011) used masked priming to explore English compound processing by adult Korean–English bilinguals. It should be noted that Ko used constituents as primes and compounds as targets, and therefore, her data may not reflect early automatic processing of compounds. Ko used a masked priming lexical decision task with both a forward mask (500ms) and a backward mask (150ms) with a prime duration of 50ms. Ko did not find a masked priming effect in any of the four conditions (+M+S+O, +M-S+O, -M-S+O, and -M+S-O). She concluded that either unbalanced adult Korean–English bilinguals do not rely on morphological decomposition in L2 processing, or her participants may not have accessed the L2 primes at all in her task condition.

The present study

The present study investigated early automatic processes involved in visual recognition of English bimorphemic compound words in native and non-native processing, using masked priming in lexical decision tasks. Chinese-speaking learners of English were chosen for the L2 group because their L1 Chinese makes extremely wide use of compounding (i.e., over 70% of words used in Modern Chinese are compounds, according to the Institute of Language Teaching and Research, 1986) and uses a different script (i.e., logographic) than English (i.e., alphabetic). The morphological, semantic and orthographic relations between prime and target words were manipulated to examine (a) whether a masked morphological priming effect can be observed in L1 and L2 compound processing, and (b) if so, whether the observed masked priming effect is mediated by semantic transparency at the early stage of processing in L1 and L2. In addition, this study examined whether constituent position plays a role in compound processing, i.e., whether a compound word primes its constituents to an equivalent extent. If it does not prime its constituents equivalently, then we want to determine which of its constituents the compound primes to a greater extent, the word-initial constituent or the word-final constituent? Finally, this study aimed to compare similarities and differences between L1 and L2 processing of English compounds.

Four experiments were conducted: Experiments 1a–b examined native processing of compound words, with Exp. 1a focusing on the priming of the word-initial compound constituents, using the compound as a masked prime and its first constituent as a target (e.g., toothbrush-TOOTH), and Exp. 1b focusing on the priming of the word-final compound constituents (e.g., toothbrush-BRUSH). Experiments 2a–b examined non-native processing of compounds, with Exp. 2a again focusing on the priming of the word-initial constituents and Exp. 2b on the priming of word-final constituents.

Participants

Fifty NSs of English (L1 group) participated in this study, 24 in Exp. 1a and 26 in Exp. 1b. They were all undergraduate students from the University of Maryland (UMD), and participated for course credit. Forty-six Chinese learners of English (L2 group) participated, 23 in Exp. 2a and 23 in Exp. 2b, and were each paid $10 for their participation. One NNS participant's data were excluded due to high error rate (>20%); accordingly, 45 L2 learners’ data remained for analysis. The 45 L2 participants (35 females) were all graduate students from China studying at UMD at the time of testing, aged 22–33 (median 23, mean 24.25). They were born in China, had been first exposed to English in formal classroom settings between 5 to 15 years old (median 10, mean 9.69), had extensive classroom learning experience in their home county (mean 12.07 years), and did not come to the U.S. or an English-speaking country until at least 18 years old (AoA, i.e., Age of Arrival, mean 22.20, range 18–30). By the time of testing, they all had been studying in the U.S. or an English speaking country for at least one year (mean 2.37 years, range 1–9 years). TOEFL iBT (Test of English as a Foreign Language Internet-Based Test) score (N = 33, mean = 101.73, range = 94–111) was taken as a measure of learners’ English proficiency for those whose length of residence (LOR) in an English-speaking country was no longer than two years. For those who had been in an English-speaking country for more than two years (N = 12), their LOR was used as an indication of L2 proficiency. All the L2 participants can be considered advanced learners of English.

Table 1 summarizes the background information of the L2 participants in Exp. 2a and Exp. 2b. The two groups were comparable in terms of age of testing, gender distribution, age of start, length of instruction, AoA, LOR, TOEFL score, and self-rated scores in listening, speaking, reading and writing.

Table 1. Background information of the Chinese learners of English in Experiments 2a and 2b

¹There is one missing value for Age and Age of Arrival, respectively in Experiment 2b; the n size therefore is 21 for them in Experiment 2b.

²TOEFL scores were collected for those whose LOR is no longer than two years. The n size for TOEFL scores in Experiment 2a is 19; the other four participants in this experiment have a mean LOR of 5 years. The n size for TOEFL scores in Experiment 2b is14; the other eight participants in this experiment have a mean LOR of 5.5 years.

³The scores for speaking, listening, reading, and writing skills are self-rated on a scale from 1 (minimal) to 10 (near-native).

Design and Stimuli

Table 2 shows the design of the study, which was a 3 (Condition/Prime Type: +M+S+O, +M-S+O, -M-S+O) × 2 (Relatedness: related, unrelated) × 2 (Target position: 1^st constituent, 2^nd constituent) design. In terms of the classification of compounds based on semantic transparency, Libben, Gibson, Yoon & Sandra (Reference Libben, Gibson, Yoon and Sandra2003)'s classification was applied for the purpose of this study: TT (Transparent-Transparent) (e.g., bedroom), OT (Opaque-Transparent) (e.g., strawberry), TO (Transparent-Opaque) (e.g., jailbird), and OO (Opaque-Opaque) (e.g., hogwash). The TT compounds were used in the +M+S+O condition, and the OO compounds were used in the +M-S+O condition.

Table 2. Design and Example Stimuli

The English Lexical Project (Balota, Yap, Cortese, Hutchison, Kessler, Loftis, Neely, Nelson, Simpson & Treiman, Reference Balota, Yap, Cortese, Hutchison, Kessler, Loftis, Neely, Nelson, Simpson and Treiman2007) was used to select the experimental stimuli based on word length, log-frequency and orthographic neighborhood size. Only compounds with noun-noun combinations (e.g., toothbrush) were used in the study. A set of 16 fully transparent compounds (TT) and 16 completely opaque compounds (OO) that had been piloted with NSs to collect subjective ratings of semantic transparency, and with Chinese learners of English in the target L2 population to check word familiarity (see details below), served as morphological primes with their constituents as targets to test word-initial (e.g., toothbrush-TOOTH) and word-final constituent priming (e.g., toothbrush-BRUSH). In the orthographic overlap condition, two sets of monomorphemic words were used as primes. One set contained words with an embedded pseudo-morpheme in word-initial position and a non-morphological ending (e.g., restaurant-REST), whereas the other set contained words with an embedded pseudo-morpheme in word-final position and a non-morphological onset (e.g., tomorrow-ROW). The former were used as controls for the testing of word-initial constituent priming and the latter for the testing of word-final constituent priming. When selecting these items, care was taken to make sure that the related orthographic prime and its target overlapped not only in orthography but also in phonology, since that was the case for the related prime-target pairs in the morphological conditions. Unrelated control compounds were selected to match the compound primes, and unrelated control primes that also contain an embedded pseudo-morpheme in word-initial or word-final position were selected to match with the orthographic-overlap primes. The Appendix contains complete lists of the experimental items for both positions.

The lexical properties of the experimental stimuli across the TT, OO, and Ortho conditions were matched with one another in terms of prime length and frequency as well as target length, frequency, and orthographic neighborhood size, both in word-initial and word-final positions (see Tables 3 and 4; all F < 1.55, all p > .22, one-way ANOVA). The unrelated control primes were paired with the related primes on length and frequency across each of the three conditions (all F < 0.20, all p > .60, one-way ANOVA). The two lists of targets (the 1^st constituents vs. the 2^nd constituents) were matched on word length, log frequency, and orthographic neighborhood size (all p > .20, t-test).

Table 3. Stimuli properties across conditions (Word-initial position / 1^st constituents as targets)

Note: the number of items in the original set for each condition (i.e., TT, OO, Orth1) is 16. After the item exclusion procedures, the number of items for each condition (i.e., TT, OO, Orth1) is 14, 12, 8, respectively, for RT analysis.

Table 4. Stimuli properties across conditions (Word-final position / 2^nd constituents as targets)

Note: the number of items in the original set for each condition (i.e., TT, OO, Orth2) is 16. After the item exclusion procedures, the number of items for each condition (i.e., TT, OO, Orth2) is 14, 12, 8, respectively, for RT analysis.

The semantic transparency of the 32 critical compounds (which served as related morphological primes) was rated by a group of 60 NSs of English. The participants were asked to rate on a 4-point scale, the extent to which the constituent morpheme (either the 1^st or the 2^nd) contributes to the overall meaning of the compound word, ranging from Not at all (1), Very little (2), to Somewhat (3), and To a great extent (4). The words were presented in pairs, with the first member of the pair always the compound and the second either its 1^st or 2^nd morpheme. Two lists were created such that the 32 compound words occur on each list only once, with half of the compounds paired with its 1^st constituent, and the other half with its 2^nd constituent. Items were randomly ordered within each list. Thirty NSs completed List A and another thirty completed List B. The results show that the opaque compounds were rated much lower than the transparent compounds with both their first constituents (M_TT = 3.54; M_OO = 2.23), t (30) = 5.408, p < .001, and their second constituents (M_TT = 3.58; M_OO = 2.19), t (30) = 8.048, p < .001.

An additional 48 words, including 32 compound words, and 16 monomorphemic words were used as primes for 48 nonword targets. The nonword targets were matched with the real word targets in length. The primes were matched with experimental primes on length and log frequency.

The targets for each position (word-initial and word-final) were divided into two counterbalanced lists such that half of the experimental targets were preceded by related primes and the other half by unrelated primes. Each participant in each experiment completed a single list. The total number of trials in each list was 96.

Results and discussion

Two NS participants’ data in Exp. 1b were excluded from analysis because they reported awareness of the existence of the primes. One NNS participant's data were excluded due to high performance error rate (>20%). Items with either the prime word or the target word for which a NNS failed to provide the correct translation were excluded from RT analysis for that participant. This procedure resulted in an exclusion of ten items (including ten related prime-target pairs and their corresponding ten unrelated control pairs; marked by an asterisk * before the item number in the Appendix) for each position, due to the high error rates for those items in the post-hoc word translation task from the NNSs. In addition, an anonymous reviewer pointed out that a few monomorphemic prime words in the orthographic overlap conditions can be exhaustively decomposed into potential morphemes, which are similar to the “corn-er” type of stimuli that have been argued to yield masked morphological priming in Rastle et al. (Reference Rastle, Davis and New2004). A close scrutiny of all items in the orthographic conditions revealed that there were four such items for each position (i.e., scar-let, log-ist-ic, earn-est, pump-kin for the word-initial position, and pro-found, inter-view, com-plain, ob-serve for the word-final position, marked by a double asterisk ** in the Appendix). An analysis on these eight “corner” type of items revealed a significant priming effect (p = .005) when the RT data from both NSs and NNSs for both positions were included. A decision was then made to exclude these items from the final analysis.

After the above exclusion procedures, the remaining 34 items, 14 for the TT condition, 12 for the OO condition, and eight for the orthographic control condition (for each position, respectively), were included in the final analysis. The lexical properties of these items across the three conditions were still matched well with each other (see Tables 3–4). The two lists of targets were also matched on word length, log frequency, and orthographic neighborhood size (all p > .63). The mean semantic transparency ratings of the remaining compounds in the TT and OO conditions for both positions were even more distant (Position 1: M_TT = 3.66; M_OO = 2.10; Position 2: M_TT = 3.58; M_OO = 1.98; both p <.001).

Only correct responses to real words were included in RT analysis. This exclusion procedure resulted in the removal of 5.3% of the RT data in Exp. 1a (NSs, Position 1), 6.7% in Exp. 1b (NSs, Position 2), 12.7% in Exp. 2a (NNSs, Position 1), and 13.1% in Exp. 2b (NNSs, Position 2). We dealt with RT outliers by establishing cut-offs at 2.5 standard deviations above and below each participant's mean RT. This resulted in a loss of 2.0% of the data in Exp. 1a and Exp. 2b, 2.1% in Exp. 1b, and 1.4% in Exp. 2a. In total, 7.2% of the RT data were removed in Exp. 1a, 8.8% in Exp. 1b, 14.1% in Exp. 2a, and 15.1% in Exp. 2b, respectively. None of the RTs included for final analysis were below 300ms or above 1500ms.

The RTs in each experiment were analyzed using linear mixed-effects models and the accuracies using generalized linear mixed-effects models in SPSS 21.0. There was no averaging of the data prior to the analyses. All RTs were inverse-transformedFootnote ¹ (i.e., -1000/RT) to reduce the positive skew in the distributions. When fitting mixed-effects models for each experiment, prime type, relatedness and their interaction were included as fixed effects, and participants and items were modeled as random variables (Baayen, Davidson & Bates, Reference Baayen, Davidson and Bates2008). We started with a basic model in which all parameters are fixed and then added random intercepts for participants and items, and then random slopes for participants and items, following Field's (Reference Field2013) recommendation (p. 831). After a series of model fitting, random intercepts for both participants and items were included in all models because their inclusion significantly improved the fit of the models using χ² likelihood ratio tests (Baayen et al., Reference Baayen, Davidson and Bates2008). Random slopes for participants and items, however, were not included because the inclusion of either or both of them did not improve the models significantly. For the purpose of the present study, we report results of F tests for each of the fixed variables (including their interaction) from the model of the best fit for each experiment and pairwise comparisons depicting the effect of relatedness (priming) at different levels of prime type (TT, OO and Orth).

In the following, we report the RT and accuracy results for each experiment in order. A joint analysis of the RT data from the four experiments will then follow, taking into consideration two more fixed variables, i.e., Position (word-initial vs. word-final), and Nativeness (native vs. non-native).

Experiment 1a. Priming of word-initial position in NSs of English

Experiment 1a focused on the priming of word-initial position in native processing of English compounds. The means and number of observations for raw RTs (after data exclusion procedures) and accuracy rates for each condition are presented in Table 5. The inverse-transformed RT data revealed a significant main effect of Relatedness, F (1, 703) = 11.096, p = .001, and of Prime Type, F (2, 33) = 4.558, p = .018; however, the Prime Type × Relatedness interaction did not reach significance, F (2, 703) = 1.680, p = .187. This pattern of results suggests that the effect of relatedness can be generalized to all levels of prime type and that the magnitude of relatedness/priming effect does not differ across prime types. It was hypothesized, however, that there should be no significant priming (or relatedness) effect in the purely orthographic overlap condition in L1 processing under such design based on previous findings (e.g., Fiorentino & Fund-Reznicek, Reference Fiorentino and Fund-Reznicek2009). In order to see whether significant priming effect existed for each of the three prime type conditions, planned comparisons regarding the effect of Relatedness at each Prime Type level were conducted. The results from such comparisons yielded a significant facilitative priming effect for the transparent condition (26 ms), F (1, 701) = 9.401, p = .002, and for the opaque condition (29 ms), F (1, 702) = 8.929, p = .003, but no significant priming for the orthographic overlap condition, F (1, 705) = 0.066, p = .798. While the interaction between Relatedness and Prime type did not reach significance, which could partially be due to the reduced number of items included in the final analysisFootnote ² and the relatively small number of participants, we have evidence from multiple comparisons that there was no priming in the purely orthographic overlap condition but there were significant priming effects in the compound conditions. The priming effects observed in the compound conditions could not simply be due to form overlap because such effect was not observed in the purely orthographic overlap condition.

Table 5. Mean reaction times and accuracy rates per condition in Exp. 1a (NSs, Position 1)

**<.01.

The role of semantic transparency was tested by examining the interaction between Relatedness (priming) and Prime Type (at TT and OO levels only) and the interaction turned out to be non-significant, F (1, 535) = .046, p = .830. This non-significant interaction suggests that the magnitude of priming did not differ across the TT and OO conditions.

The accuracy analysis showed no significant main effect of either Prime Type, F (2, 810) = 2.347, p = .100, or Relatedness, F (1, 810) = 0.192, p = .662. The interaction between Prime Type and Relatedness was not significant either, F (2, 810) = 0.340, p = .832.

In Exp. 1a, significant and statistically equivalent facilitative masked priming effects were obtained for the word-initial constituents of semantically transparent and opaque compounds while no priming was observed in the orthographic overlap condition. These results replicate the findings of Experiment I in Fiorentino and Fund-Reznicek (Reference Fiorentino and Fund-Reznicek2009), the earlier study that examined native processing of English compounds using the same task and design. This pattern of priming effects also converges with the majority of those observed for the root of derivationally suffixed words under similar masked priming conditions (e.g., Longtin, Segui & Hallé, Reference Longtin, Segui and Hallé2003; Rastle et al., Reference Rastle, Davis, Marslen-Wilson and Tyler2000; Rastle et al., Reference Rastle, Davis and New2004). This pattern contrasts with the findings from Diependaele et al. (Reference Diependaele, Sandra and Grainger2005) and Diependaele et al. (Reference Diependaele, Duñabeitia, Morris and Keuleers2011) that examined French and English suffixed derivations, in that significantly larger facilitative priming was found for semantically transparent items relative to opaque items in their studies, but no such semantic transparency effect was observed in this current experiment. The findings of this experiment thus contribute evidence in favor of a fast automatic sublexical morpho-orthographic decomposition mechanism independent of semantic transparency in native compound processing.

Experiment 1b. Priming of word-final position in NSs of English

Experiment 1b examined the priming of word-final position in native processing of English compounds. The raw mean RTs and accuracy rates for each condition are presented in Table 6. The inverse-transformed RT data revealed a significant main effect of Relatedness, F (1, 688) = 5.812, p = .016. The main effect of Prime Type did not reach significance, F (2, 32) = 2.463, p = .101. The interaction between Prime Type and Relatedness was marginally significant, F (2, 689) = 2.702, p = .068. Planned comparisons regarding the effect of Relatedness at each Prime Type level revealed a significant facilitative priming effect for the TT condition (22 ms), F (1, 690) = 5.162, p = .023, and for the OO condition (21 ms), F (1, 688) = 8.918, p = .003, but no significant priming for the Orth condition, F (1, 688) = 0.281, p = .596.

Table 6. Mean reaction times and accuracy rates per condition in Exp. 1b (NSs, Position 2)

*<.05;

**<.01.

As for the role of semantic transparency, the interaction between Relatedness (priming) and Prime Type (at TT and OO levels only) was not significant, F (1, 526) = .343, p = .558, suggesting that the magnitude of priming did not differ across the TT and OO conditions.

The accuracy analysis showed no significant main effect of either Prime Type, F (2, 810) = 2.207, p = .111, or Relatedness, F (1, 810) = 2.818, p = .094. The interaction between Prime Type and Relatedness was not significant either, F (2, 810) = 0.275, p = .759.

In this experiment, significant and statistically equivalent priming effects were obtained for the word-final constituents of semantically transparent and opaque compounds, while no priming was observed in the orthographic control condition. These findings replicate those in Experiment 1a as well as those in Fiorentino and Fund-Reznicek (Reference Fiorentino and Fund-Reznicek2009), Experiment II, that examined word-final constituent masked priming in native processing. This pattern of priming effects for the word-final constituents of compounds also converges with the pattern of priming observed for the root of derivationally prefixed words in Dutch reported by Diependaele et al. (Reference Diependaele, Sandra and Grainger2009), Experiment 1, under similar masked priming conditions in L1 processing. The same patterns of priming effects in Experiments 1a and 1b in the current study provide converging evidence for a fast automatic sublexical morpho-orthographic decomposition mechanism independent of semantic transparency in native processing of compounds.

Experiment 2a. Priming of word-initial position in Chinese learners of English

Experiment 2a explored the priming of word-initial constituents in non-native processing of English compounds by a group of advanced Chinese learners of English. The raw mean RTs and accuracy rates for each condition are presented in Table 7. The analysis of the inverse-transformed RT data revealed a significant main effect of Relatedness, F (1, 622) = 36.356, p < .001. The main effect of Prime Type was not significant, F (2, 33) = 1.049, p = .362, nor was the interaction between Prime Type and Relatedness, F (2, 622) = 1.011, p = .365. Planned comparisons evaluating the effect of Relatedness at each Prime Type level yielded significant facilitative priming effects for the TT condition (18 ms), F (1, 618) = 10.341, p = .001, the OO condition (65 ms), F (1, 620) = 22.695, p < .001, as well as for the Orth condition (74 ms), F (1, 625) = 7.891, p = .005.

Table 7. Mean reaction times and accuracy rates per condition in Exp. 2a (NNSs, Position 1)

**<.01.

The role of semantic transparency was again tested by examining the interaction between Relatedness and Prime Type (at TT and OO levels only). The interaction failed to reach significance, F (1, 493) = 1.776, p = .183, indicating that the magnitude of priming effects did not differ significantly across the TT and OO conditions.

As for accuracy analysis, unlike NSs who demonstrated no effect for either Relatedness or Prime Type, advanced Chinese learners of English showed a significant main effect of Prime Type, F (2, 776) = 18.124, p < .001, and of Relatedness, F (1, 776) = 9.764, p = .002. The interaction between Prime Type and Relatedness was not significant, F (2, 776) = 0.373, p = .689. Accuracy rates were significantly higher in the TT and OO conditions (with no significant difference between them) than that in the Orth condition, and accuracy following related primes was higher than following unrelated control primes.

Experiment 2a, which tested the priming of word-initial constituents in non-native processing of compounds, elicited a significant masked priming effect not only in the transparent and opaque compound conditions, but also in the orthographic overlap condition. Similar to L1 processing, significant facilitative masked priming effects were observed in L2 processing for both transparent and opaque compound primes, and the magnitudes of the priming effects did not differ significantly between the TT and OO conditions. Unlike NSs, however, Chinese-speaking learners showed a clear facilitative form priming effect when the target word overlaps with the initial part of the prime. This finding will be discussed in more detail in the subsequent general discussion section.

Experiment 2b. Priming of word-final position in Chinese learners of English

Experiment 2b investigated the priming of word-final constituents in non-native processing of English compounds by another group of advanced Chinese learners of English. The raw mean RTs and accuracy rates for each condition are presented in Table 8. The analysis of the inverse-transformed RT data showed a significant main effect of Relatedness, F (1, 582) = 6.136, p = .014, but no significant main effect of Prime Type, F (2, 33) = .244, p = .785. More importantly, a significant interaction between Prime Type and Relatedness was found, F (2, 583) = 3.909, p = .021, suggesting that the effect of relatedness changes depending on the level of prime type. Planned comparisons gauging the effect of Relatedness at each Prime Type level yielded significant facilitative priming effects for the TT condition (32 ms), F (1, 584) = 9.325, p = .002, and for the OO condition (19 ms), F (1, 581) = 8.223, p = .004, while no significant facilitation was found in the Orth condition, F (1, 582) = .813, p = .368.

Table 8. Mean reaction times and accuracy rates per condition in Exp. 2b (NNSs, Position 2)

**<.01.

In terms of the role of semantic transparency, the interaction between Relatedness and Prime Type (at TT and OO levels only) was not significant, F (1, 446) = .011, p = .917, indicating that the magnitude of priming effect did not differ significantly for transparent and opaque primes.

The accuracy results showed a significant main effect of Relatedness, F (1, 742) = 4.881, p = .027, with higher accuracy following related primes. The main effect of Prime Type was not significant, F (2, 742) = 1.276, p = .280, nor was the interaction between Prime Type and Relatedness, F (2, 742) = 0.125, p = .883.

Experiment 2b, which tested the priming of word-final constituents in L2 processing of compounds, elicited significant and statistically equivalent masked priming in the transparent and opaque compound conditions and no priming in the orthographic overlap condition. This pattern of priming effects converges with those observed in Experiments 1a and 1b as well as those reported in Fiorentino and Fund-Reznicek (Reference Fiorentino and Fund-Reznicek2009) on L1 processing of English compounds. These results seem to suggest that advanced Chinese learners of English use the same morphological processing strategy as native speakers. That is, an automatic sublexical morpho-orthographic decomposition mechanism that is independent of semantic transparency also seems to operate in advanced Chinese learners of English.

Joint Analysis

We merged the RT data from the four experiments to explore the effects of two more variables, i.e., Position (word-initial vs. word-final) and Nativeness (native vs. non-native), in addition to Prime Type and Relatedness. A significant main effect was found for Nativeness, F (1, 94) = 63.568, p < .001, with NSs responding significantly faster than NNSs, and for Relatedness, F (1, 2657) = 50.951, p < .001, with RTs significantly shorter following related primes. The main effect of Prime Type was marginally significant, F (2, 66) = 3.004, p = .056. The main effect of Position was not significant, F (1, 142) = 0.392, p = .532. The four-way interaction between Nativeness, Position, Prime Type, and Relatedness was not significant, F (2, 2657) = 1.031, p = .357, nor were any of the three-way interactions (all p > .113).

The two-way interaction of Prime Type and Relatedness was significant in this big model, F (2, 2658) = 5.926, p = .003, with the Relatedness effect significant at the transparent and opaque prime type levels (23 ms for the transparent prime type, F (1, 1066) = 33.096, p < .001; 32 ms for the opaque prime type, F (1, 894) = 42.049, p < .001), but nonsignificant at the orthographic control level (F (1, 537) = 0.759, p = .384). It is important to note that although the Prime Type × Relatedness interaction did not reach statistical significance in each of the four separate models in the four experiments (i.e., it was significant in L2 processing at Position 2 (p = .021), marginally significant in L1 processing at Position 2 (p = .068), and not significant at Position 1 for both the L1 group (p = .187) and the L2 group (p = .365)), the interaction did turn out to be significant in the joint analysis model, providing support to our interpretation that the compound priming effect observed in the study could not be simply due to orthographic overlap.

The two-way interaction between Position and Relatedness was also significant, F (2, 2657) = 5.427, p = .020. The effect of Relatedness was significantly larger at the word-initial overlap position (30 ms, F (1, 1351) = 48.844, p < .001) than at the word-final overlap position (17 ms, F (1, 1301) = 19.663, p < .001). Although the three-way interaction between Nativeness, Position and Relatedness did not reach statistical significance, F (1, 2654) = 2.520, p = .113, separate analyses for the two language groups showed that the interaction of Position and Relatedness was not significant for the English NS group, F (1, 1390) = .337, p = .562, and was only significant for the NNS Chinese group, F (1, 1201) = 5.579, p = .018. These results suggest that there was no position effect in L1 processing, but there was a position effect in L2 processing with larger priming effect for the word-initial overlap position. Relating this to the results of the four separate models presented earlier (see Tables 5–8), we suspect that the larger priming effect for the word-initial position in L2 processing is likely to be driven by the presence of the large priming effect in the orthographic control condition at that position. We then decided to run another joint analysis of the RT data from the four experiments, excluding those in the orthographic control conditions, to clarify any effects of Position and Nativeness in compound processing.

When the orthographic control conditions were excluded from the joint analysis, there was again a significant main effect for Relatedness, F (1, 2040) = 79.208, p < .001, and for Nativeness, F (1, 93) = 62.426, p < .001. The RTs following related primes were significantly shorter than those following unrelated primes in the compound conditions. In addition, native English speakers were 114 ms faster than Chinese learners of English across the board in compound processing (NS Grand Mean RT = 562 ms; NNS Grand Mean RT = 676 ms). Other than these two main effects, none of the other main effects (including Position, F (1, 130) = .451, p = .503 and Prime Type, F (1, 50) = .274, p = .603) were significant, nor were any of the two-way, three-way or four-way interactions. The interaction of Prime Type (TT & OO) and Relatedness was not significant, F (1, 2049) = 1.229, p = .268. The interaction of Position and Relatedness was not significant either, F (1, 2030) = 1.130, p = .288. The interaction of Nativeness × Position × Relatedness was not significant, F (1, 2040) = .445, p = .505. The four-way interaction of Nativeness × Position × Prime Type × Relatedness was not significant, F (1, 2049) = .616, p = .433. These results suggest that the magnitudes of the priming effects for the transparent and opaque compounds did not differ across positions in either NS or NNS processing of compounds (i.e., no Position effect) and that except for being slower, this group of advanced Chinese learners of English process English compound words in similar ways as NSs of English (i.e., no Nativeness effect).