First language (L1) acquisition from a usage-based perspective

A much-debated issue in L1 acquisition pertains to how compositional phrases (those with a meaning derivable from word parts; e.g., don’t have to worry, as opposed to sequences like He left the US for good, where for good cannot be taken apart to derive its meaning) are acquired and processed. The mainstream generative approach to language (Chomsky, Reference Chomsky1995; Pinker, Reference Pinker1999; Prasada & Pinker, Reference Prasada and Pinker1993) argues for a human innate language-specific acquisition device and a separation of the lexicon and the grammar in a speaker’s linguistic representation. Moreover, frequency effects, a psychological characteristic of memory, should be observed only from memorized items in the mental lexicon, not from processing of compositional phrases, which are generated from abstract grammar rules. By contrast, in usage-based approaches, language acquisition results from the interaction between input and human domain-general cognitive processes, and the lexicon and the grammar are not rigidly divided (e.g., Bybee, Reference Bybee2010; Tomasello, Reference Tomasello and Bavin2009). Human linguistic knowledge consists of constructions (e.g., the English transitive) or form–meaning mappings in a particular language (e.g., subject–verb–object with the prototypical meaning of one entity acting on the other), ranging from morphemes, words, idioms, and partially lexically filled linguistic patterns (e.g., The __er the __er) to fully general and abstract linguistic patterns (e.g., subject–verb–object; Goldberg, Reference Goldberg2003, Reference Goldberg2006). Moreover, constructions with varying sizes and degrees of abstraction are represented and processed based on similar general mechanisms, and speakers can simultaneously retain the representation of both a more general and abstract construction (e.g., subject–verb–object) and more lexically specific related forms (e.g., David kicked the ball, KICKER kicked OBJECT; Ambridge & Lieven, Reference Ambridge and Lieven2011).

From a usage-based view, L1 construction acquisition results from an accumulation of statistical probabilities and abstraction of regularities out of previous construction encounters. As Ellis (Reference Ellis1996, Reference Ellis2002) noted, frequency, a key index of statistical information, promotes entrenchment of words, and the cognitive ability of chunking allows strings of words to be registered in human memory. Subsequent exposure to massive input results in statistical fine-tuning, making the registered relation reflect frequency of encounters. Usage-based theories posit that frequency affects acquisition of constructions at all levels. Thus, L1 speakers should exhibit sensitivity to frequency of not only single words but also compositional phrases. That is, in language comprehension and production, more frequent compositional phrases should be recognized, comprehended, and produced faster (e.g., Ambridge, Kidd, Rowland, & Theakston, Reference Ambridge, Kidd, Rowland and Theakston2015; Arnon, Reference Arnon2015; Bybee, Reference Bybee2010).

Despite the similar views regarding the general L1 acquisition mechanism, as Arnon and Snider (Reference Arnon and Snider2010) pointed out, there are two different assumptions in usage-based approaches regarding compositional phrase representation. One assumption, grounded in work of researchers such as Goldberg (Reference Goldberg2006) and Wray (Reference Wray2002), is that “phrases that are of sufficient frequency can attain independent representation as a way of making processing more efficient” (Arnon & Snider, Reference Arnon and Snider2010, p. 69). There is not yet a clear consensus as to the minimum frequency level (i.e., sufficient frequency), but in this view, there is a qualitative difference between highly frequent phrases (i.e., stored as a whole) and less frequent phrases (i.e., generated or analyzed by the language grammar), and the first type is processed faster than the second. The other assumption, informed by work of researchers such as Bybee and colleagues (Bybee, Reference Bybee2006, Reference Bybee2010; Bybee & Hopper, Reference Bybee and Hopper2001), is that there is no such qualitative difference. Speakers retain information about previously encountered phrases, and more frequent phrases are more entrenched in speakers’ representation. The difference between a more frequent phrase and a less frequent phrase is quantitative (i.e., a difference in the level of activation), resulting from different frequencies of previous phrase encounters. Therefore, phrases need not be highly frequent to be processed faster; relatively more frequent phrases should be processed faster than less frequent phrases regardless of the frequency range. Despite these two different assumptions, compositional phrase frequency effects are accounted for in usage-based approaches.

Previous work on phrase frequency effects in native English speakers (NSs)

Numerous empirical studies have corroborated the ontological status of compositional phrase frequency effects during receptive language processing in NSs (e.g., Arnon & Snider, Reference Arnon and Snider2010; Reali & Christiansen, Reference Reali and Christiansen2007; Siyanova–Chanturia, Conklin, & van Heuven, Reference Siyanova-Chanturia, Conklin and van Heuven2011; Sosa & Macfarlane, Reference Sosa and MacFarlane2002; Tremblay, Derwing, Libben, & Westbury, Reference Tremblay, Derwing, Libben and Westbury2011). Among these, studies by Arnon and Snider (Reference Arnon and Snider2010) and Hernández, Costa, and Arnon (Reference Hernández, Costa and Arnon2016) have provided relatively strong support as indicated by reaction times in a phrasal acceptability task because these two studies used phrases longer than two words (i.e., don’t have to worry) and controlled for frequencies of subparts of target phrases (e.g., don’t have). Such a control is important; without it, frequency of the whole target phrases may not have been the only factor determining the results.

Regarding production, studies with NSs have reported phrase frequency effects on phonetic reductions and voice onset time (e.g., Aylett & Turk, Reference Aylett and Turk2004; Jannsen & Barber, Reference Jannsen and Barber2012). Some other studies investigated the role of frequency on production durations of compositional phrases longer than two words in elicited or spontaneous speech. The results have been mixed. First, using an elicited production task, Bannard and Matthew (Reference Bannard and Matthews2008) and Arnon and Priva (Reference Arnon and Priva2013) reported frequency effects in children and adults, respectively, after frequency of the subparts of the target phrases was controlled for. However, using a similar task, Ellis, Simpson-Vlach, and Maynard (Reference Ellis, Simpson-Vlach and Maynard2008) and Tremblay and Tucker (Reference Tremblay and Tucker2011) observed no effects. A possible cause for the lack of the effects in these latter two studies may have been methodological. For example, in Ellis et al.’s (Reference Ellis, Simpson-Vlach and Maynard2008) study, there was a lack of subpart frequency control, and while some of the stimuli were complete syntactic constituents (e.g., a great deal of), others were not (e.g., and at the). Regarding spontaneous phrase production durations, there is evidence for compositional phrase frequency effects based on telephone conversations (Arnon & Priva, Reference Arnon and Priva2013) and interview speech (Arnon & Priva, Reference Arnon and Priva2014) when subpart frequencies were accounted for. Collectively, therefore, research on L1 speech production durations has lent some support to frequency effects on production of compositional phrases longer than two words, but cross-study methodological differences (e.g., type of speech analyzed and substring frequency control) may have been responsible for the inconclusive results.

Frequency in second language (L2) acquisition

Recent years have seen an increase in usage-based L2 research (e.g., Ellis, Reference Ellis2019; Ellis, Römer, & O’Donnell, Reference Ellis, Römer and O’Donnell2016; Ortega, Reference Ortega2013; Supasiraprapa, Reference Supasiraprapa2018) due to the proposal that, as in L1 acquisition, L2 acquisition is driven by accumulation of statistical probabilities of previously encountered input (e.g., Ellis, Reference Ellis, Gass and Mackey2011, Reference Ellis, Trousdale and Hoffmann2013). However, the L2 literature has also documented various L1–L2 acquisition differences, which may lead to different frequency effects between adult NSs and adult non-native English speakers (NNSs). First, the input that L1 learners and adult L2 learners receive may differ in terms of amount and structure (e.g., Littlemore, Reference Littlemore2009; Muñoz, Reference Muñoz2008). Second, while phrase frequency effects result from implicit learning mechanisms (e.g., Ellis, Reference Ellis2002, Reference Ellis, Trousdale and Hoffmann2013) and while L1 and L2 abstract learning patterns may share some similarities (e.g., Wonnacott, Reference Wonnacott2011), compared to L1 learners, older L2 learners may be less apt at acquiring linguistic patterns implicitly and rely more on explicit learning (e.g., Bley-Vroman, Reference Bley-Vroman2009; DeKeyser, Reference DeKeyser2000). As Wray (Reference Wray2002) observed, L2 learners may also not retain memory about L2 word co-occurrences but instead break phrases into individual words due to various reasons: a lack of necessity to memorize and use L2 word phrases, the typical focus on individual words in adult L2 education, and their mature cognitive development and L1 literacy, which prompt them to break lexical sequences into words. Research on frequency effects on L2 compositional phrase processing is thus interesting because it can potentially shed light on L1–L2 acquisition similarities and differences.

Previous work on frequency effects in L2 English phrase comprehension and production

To date, empirical research on receptive L2 English compositional phrase processing has lent some support to phrase frequency effects. For example, three studies (Gyllstad & Wolter, Reference Gyllstad and Wolter2016; Wolter & Gyllstad, Reference Wolter and Gyllstad2013; Wolter & Yamashita, Reference Wolter and Yamashita2018) found that, in phrasal decision tasks, adult NNSs processed English collocations faster when collocation frequency increased. In addition, a study by Sonbul (Reference Sonbul2015), which used a reading task with concurrent eye-movement registration, reported that relatively proficient adult NNSs demonstrated sensitivity to frequency of English collocations as measured by first pass reading time. In these studies, phrase frequency effects were importantly documented when frequencies of the subparts of the target collocations were controlled for. However, the stimuli consisted of two words (e.g., middle class and pay tax); as in L1 acquisition, stronger evidence for phrase frequency effects would be from processing of longer phrases. Moreover, in the case of Sonbul’s (Reference Sonbul2015) study, participants had mixed L1 backgrounds; thus, the influence of L1 on L2 collocation processing (Wolter & Gyllstad, Reference Wolter and Gyllstad2011, Reference Wolter and Gyllstad2013; Wolter & Yamashita, Reference Wolter and Yamashita2018; Yamashita & Jiang, Reference Yamashita and Jiang2010) was not accounted for.

Three other studies investigated compositional three- to five-word sequence processing, offering support (Ellis et al., Reference Ellis, Simpson-Vlach and Maynard2008; Siyanova-Chanturia et al., Reference Siyanova-Chanturia, Conklin and van Heuven2011) or a lack thereof (Valsecchi et al., Reference Valsecchi, Kuänstler, Saage, White, Mukherjee and Gegenfurtner2013) based on adult L2 comprehension. These three studies, however, shared the same methodological limitation: a lack of subpart frequency control. Addressing this limitation, Hernández et al. (Reference Hernández, Costa and Arnon2016) more recently demonstrated with a phrasal decision task that adult L2 learners’ processing of four-word phrases was frequency sensitive, offering relatively strong support for usage-based L2 acquisition.

In terms of frequency effects on L2 English production, Durrant and Schmiit (Reference Durrant and Schmitt2010) reported that NNSs’ memory retention of collocations, as measured with a productive recall task, increased as a function of the number of times the collocations appeared in a previous training session. In addition, Fernández and Schmitt (Reference Fernández and Schmitt2015) reported a significant correlation between NNSs’ active recall of English collocations and collocation frequency. While these findings are compatible with a usage-based account, the two studies contained certain methodological limitations. For example, using only an immediate recall task, the former study focused on initial memory retention of L2 word co-occurrences, but not on whether subsequent frequency of collocation encounters (i.e., statistical fine-tuning) leads to frequency effects (Ellis, Reference Ellis1996, Reference Ellis2002). Moreover, in the latter study, frequencies of constituent words of the target collocations were not controlled for, and both studies focused on only two-word L2 collocations.

As for frequency effects on production durations of L2 English phrases longer than two words, relevant research has been scarce. In fact, it seems that there has been only one empirical study, namely, Ellis et al. (Reference Ellis, Simpson-Vlach and Maynard2008). Focusing on academic word sequences (e.g., it should be noted that), this study observed no frequency effects in both NSs and NNSs in an elicited speech production task. However, as discussed, methodological issues in this study raised a question of whether phrase frequency alone determined participants’ production durations. Research on frequency effects on NNSs’ compositional phrase production durations thus deserves further investigation.

Perhaps two additional methodological issues in the existing NNS research should be pointed out. First, in studies reporting compositional phrase frequency effects, the typical frequency data source was a large native speaker corpus, including the Corpus of Contemporary American English (Davies, Reference Davies2013), which was used by researchers such as Wolter and Gyllstad (Reference Wolter and Gyllstad2013) and Wolter and Yamashita (Reference Wolter and Yamashita2018), and the Fisher (Godfrey, Holliman, & McDaniel, Reference Godfrey, Holliman and McDaniel1992) and Switchboard (Cieri, Miller, & Walker, Reference Cieri, Miller and Walker2004) corpora of American English telephone conversations, which were used by Hernández et al. (Reference Hernández, Costa and Arnon2016). While the input that NSs and adult NNSs receive are unlikely to be identical (e.g., Muñoz, Reference Muñoz2008), such findings seemed to corroborate the proposal that data from a large and adequately representative corpus should represent the shared regularities of input that all language users have received (Hoey, Reference Hoey2005). Second, most NNS studies reporting phrase frequency effects obtained frequency data from both spoken and written native speaker corpora (e.g., Wolter & Gyllstad, Reference Wolter and Gyllstad2013), suggesting that frequency of encounters in written input may affect L2 phrase representation. To date, however, a direct discussion about the role of written texts when compared to the role of speech input in usage-based L2 acquisition has been relatively limited, possibly because L2 acquisition theories originate from child L1 acquisition theories (e.g., Ellis, Reference Ellis, Trousdale and Hoffmann2013; Tomasello, Reference Tomasello and Bavin2009). Consequently, at this point, frequency of previous encounters of compositional phrases in spoken L2 input seems most theoretically relevant.

In sum, empirical studies have offered some evidence for L1 and L2 compositional phrase frequency effects, but limitations in the existing studies included a lack of subpart frequency control and scant attention to L2 phrase production. Therefore, addressing these limitations in the following two respective experiments, the present study sought evidence for frequency effects in both receptive and productive processing of L1 and L2 compositional four-word phrases. The current study also constituted the first attempt to investigate such effects in both language comprehension and production in the same adult English L1 and L2 speakers. The research questions were as follows:

1. 1. Are adult NSs and NNSs sensitive to the frequency of compositional four-word phrases in recognition when the frequency of the smaller parts is controlled for?

2. 2. Are adult NSs and NNSs sensitive to the frequency of compositional four-word phrases during language production when the frequency of the smaller parts is controlled for?

Experiment 1

Method

Participants

Participants were 51 adult NSs (mean age = 20.58, SD = 2.93) and 52 adult NNSs (mean age = 23.54, SD = 3.93) who were undergraduate or graduate students at a large Midwestern US university. They had no record of hearing impartment or speech difficulty. The NNSs were international students from China and shared the same L1 (Chinese). They can be characterized as being proficient enough to study in an English-speaking environment (minimum internet-based Test of English as a Foreign Language score of 85, mean score = 95.52, SD = 6.63). None had an English immersion experience before the age of 10. Moreover, they had a relatively similar length of stay in the United States (range = 2–3 years, mean = 2.61, SD = 0.56) and had never lived in another English-speaking country before coming to the United States, except for 1 participant who had stayed in Canada for 5 months.

Materials

Target phrases

This experiment used the phrases from Arnon and Snider (Reference Arnon and Snider2010) that yielded phrase frequency effects in adult NSs. The target phrases were 28 pairs of phrases, each consisting of two four-word phrases that differed in phrase-frequency (high vs. low) and differed only in the final word (e.g., don’t have to worry vs. don’t have to wait). The two variants in each pair were of the same constituent type (e.g., verb phrases or noun phrases). These target pairs were constructed from the Fisher (Godfrey et al., Reference Godfrey, Holliman and McDaniel1992) and Switchboard (Cieri et al., Reference Cieri, Miller and Walker2004) corpora of American English telephone conversations, which together contained around 20 million words. The phrases are therefore common in general English conversations. At the end of Experiment 2 in the current study, when asked to identify any component words they did not know, participants who were NNSs indicated that they knew all the words in the target phrases.

As in Arnon and Snider’s (Reference Arnon and Snider2010) study, the 28 target pairs were from a high- and a low-cutoff bin, which consisted of 16 and 12 target pairs, respectively. In the high cutoff bin, based on the frequencies derived from Fisher and Switchboard, each high-frequency variant occurred at least 12.00 times per million words, while each low-frequency variant occurred less often. By contrast, in the low cutoff bin, in each pair, the high-frequency variant occurred at least 1.00 time per million words, while the low-frequency variant occurred less frequently. The purpose of incorporating the two cutoff bins was twofold. First, Arnon and Snider (Reference Arnon and Snider2010) observed frequency effects from adult NSs in a phrasal acceptability task in both bins. In light of the two previously discussed proposals about usage-based compositional phrase representation, Arnon and Snider (Reference Arnon and Snider2010) argued that phrases do not need to be highly frequent (i.e., have frequency above the cutoff point in the high cutoff bin) to be processed faster. Therefore, there was no direct evidence that highly frequent phrases were qualitatively different from phrases with lower frequencies. Instead, because frequency effects were similarly observed in both cutoff bins, the researchers argued that the differences in reaction times to the target phrases should have resulted from relative quantitative differences (i.e., different frequencies of previous phrase encounters), regardless of the frequency range (Bybee, Reference Bybee2006, Reference Bybee2010; Bybee & Hopper, Reference Bybee and Hopper2001). The current experiment therefore investigated whether Arnon and Snider’s (Reference Arnon and Snider2010) results are replicable in adult NSs. The second reason was specifically related to the NNSs, who presumably had had less exposure to English input. These learners may exhibit frequency effects only with target phrases in the high cutoff bin because they may have been exposed to many instances of highly frequent phrases. In contrast, the NNSs may not have had much exposure to the target phrases in the low cutoff bin and thus may not demonstrate frequency effects in this bin. In short, an additional goal of having the two cutoff bins was to illuminate whether adult NNSs had stored sufficient accumulated statistics information (i.e., frequency of occurrences) through previous L2 exposure to demonstrate frequency effects in both cutoff bins.

Table 1 shows examples of the target phrases. The first two pairs are examples from the high cutoff bin. The frequency cutoff point for classifying a phrase as a high- or low-frequency phrase in this bin (12.00 times per million words) was slightly higher than that in Arnon and Snider’s (Reference Arnon and Snider2010) study (10.00 times per million words) because in the current study the frequencies of the target phrases, although obtained from the same corpora, were found to be slightly higher than the frequencies reported in the original study. The cutoff point was therefore slightly increased so that all the low-frequency phrases in the original study were still considered low-frequency phrases in the current study (e.g., in the current study I don’t know who, a low-frequency phrase, was found to occur 11.60 times per million words). The third and fourth pairs are examples from the low cutoff bin, in which the cutoff point for classifying a phrase as a high- or a low-frequency phrase (1.00 time per million words) is the same as that in Arnon and Snider’s (Reference Arnon and Snider2010) study. Thus, the classification of a phrase as having high or low frequency was meaningful within each cutoff bin, not across all the 28 stimuli pairs. The 28 target pairs and their frequencies are listed in Appendix A.

Table 1. Examples of the 28 target pairs

In the high cutoff bin and based on the frequency data derived from Fisher and Switchboard, the mean frequencies of the high- and low-frequency variants across the 16 target pairs were 25 occurrences (Min = 12.00, Max = 53.15, SD = 12.97) and 4.87 occurrences (Min = 0.70, Max = 11.60, SD = 3.93) per million words, respectively. Across these pairs, frequencies between high- and low-frequency phrases differed significantly, t (30) = –10.76, p < .001. In the low cutoff bin, the mean frequencies of the high- and low-frequency variants across the 12 target pairs were 4.68 occurrences (Min = 1.85, Max = 12.60, SD = 3.18) and 0.27 occurrences (Min = 0.05, Max = 0.55, SD = 0.14) per million words, respectively. Across these 12 pairs, frequencies between high- and low-frequency phrases also differed significantly, t (22) = –4.81, p < .001. Therefore, in each cutoff bin, high-frequency phrases occurred significantly more often than low-frequency phrases.

Regarding the subparts of the phrases, each pair (e.g., don’t have to worry vs. don’t have to wait) differed only in three subparts: the last word (e.g., worry vs. wait), the last two words (e.g., to worry vs. to wait), and the last three words (e.g., have to worry vs. have to wait). To observe effects of the four-word sequence more clearly, the frequencies of these subparts will later be entered as control variables in the analysis. Finally, the high- and low-frequency phrases in each cutoff bin did not differ in terms of the plausibility of meaning. This was attested by results from a plausibility rating task completed by a group of NSs in Arnon and Snider’s (Reference Arnon and Snider2010) study. This is not surprising as all the phrases are possible and meaningful in English.

Fillers

Besides the 56 target phrases, there were 80 four-word-phrase fillers of two types. As in Arnon and Snider’s (Reference Arnon and Snider2010) study, the first were 12 possible phrases (e.g., buy a new dress), and the second were 68 impossible phrases. Among the latter, 75% had a wrong word order (e.g., she a has boat), while 25% were impossible due to an incorrect preposition use (e.g., afraid to the dark). An attempt was made to avoid an overlap between words in the target phrases and the fillers. In total, there was an equal number of possible phrases (54 target phrases plus 12 fillers) and impossible phrases (68 fillers).Footnote ¹

Procedure

The stimuli were divided into two blocks: A and B. One variant from each of the 28 target pairs was randomly assigned to only one block to minimize a possible repetition effect resulting from participants seeing the identical first three words in the two variants of the same target pair in the same block. Each block thus consisted of 14 high-frequency variants and 14 low-frequency variants from the target pairs. Fillers were randomly and equally assigned to each block such that, in total, each block contained 6 fillers, which are possible English word sequences, and 34 fillers, which are impossible sequences. The total number of phrases in each block was thus 68, with half of the phrases being grammatical and the other half ungrammatical. The stimuli were presented in a random order.

The experiment was run on Superlab (Cedrus Corporation, 2006). Only informed that the current study investigates English phrase comprehension and production, each participant sat in a quiet room in front of a computer screen and completed a phrasal decision task, in which they saw four-word phrases in the center of the screen, one at a time, and were asked to press a YES button or a NO button, which were equally positioned on a keypad, to judge whether the phrases were possible English word sequences. Two different instructions and keypads were used with right-handed and left-handed participants. The participants were instructed to make their judgment as fast as they could while still being accurate. Participants also completed a short practice section, in which they saw examples of both possible and impossible sequences, before the actual experiment began. During the experiment, participants first saw a plus sign in the center of the screen for eye fixation. The sign lasted 333 ms and was followed by a long blank screen presented for 50 ms. A phrase then appeared and remained on the screen until a button press. The phrases appeared one at a time and in their entirety (font: Arial; size: 36; position: center). Words were in the lowercase except the first-person personal pronoun and proper names.

This experiment followed a within-subject counterbalanced design. Half of the participants in each group were randomly assigned to complete Block A first, while the other half completed Block B first. Each participant completed both blocks and thus saw the two variants from each target pair across the two blocks. Between the two blocks, there was a break during which participants completed the first part of their background questionnaire. The break was included to further reduce possible repetition effects. The whole experiment took approximately 20 minutes.

Analysis

Because the frequency cutoff point in each cutoff bin differed, a separate analysis for each cutoff bin was conducted. Only reaction times for the target phrases were analyzed. One NS and 1 NNS were excluded due to relatively low levels of judgment accuracy of 88% and 77%, respectively; the minimum accuracy levels to exclude participants in the two respective groups were 93% and 89%. Excluded as well were 2 NNSs who occasionally stopped during the experiment to do unrelated activities. The remaining 50 NSs and 49 NNSs had a mean accuracy rate of 98% (SD = 0.02) and 97% (SD = 0.03), respectively. Therefore, both groups did not seem to have any difficulty doing the task. Incorrect responses were also excluded. Moreover, in each group, reaction times exceeding +/–2 SD from the group mean in each frequency condition in each cutoff bin were removed, resulting in a removal of 3% and 4% of the correct responses from the NSs and the NNSs, respectively.

The data were analyzed with mixed-effects regression models (Baayen, Davidson, & Bates, Reference Baayen, Davidson and Bates2008; Bates, Reference Bates2010). Models were run in R (R Core Team, 2015) with the statistics package lme4 (Bates, Reference Bates2010; Bates, Mächler, Bolker, & Walker, Reference Bates, Mächler, Bolker and Walker2015). Reaction times were log transformed to reduce skewness of the data.Footnote ² Fixed effects included phrase frequency condition (high/low), participant group (NS/NNS), the interaction between these two variables, and control variables: block order (i.e., whether the participant completed that block as the first or the second block), the number of characters of the target phrases, and the frequencies of the subparts that differ in each pair (e.g., frequencies of to worry and to wait in the pair don’t have to worry and don’t have to wait). The interaction between block order and phrase frequency was also included. Continuous predictors were standardized to illustrate the magnitude of such predictors. Moreover, participant and item were included in the model as random intercepts. Wald statistics (Type II) were used to determine the significance of the effects (Fox, Reference Fox2008; Fox & Weisberg, Reference Fox and Weisberg2011).

In the analysis, model comparisons suggested that adding different reaction time slopes for participant and item random effects did not significantly improve the models, either for the high cutoff bin (p = .179) or the low cutoff bin (p = .185). Consequently, the final model for each cutoff bin allowed for a random intercept for each participant and each test item, but a by-participant random slope for phrase frequency was not included. In each final model, the variance inflation factor (VIF) scores for the continuous predictors were well below the threshold level of 10 (Loewen & Plonsky, Reference Loewen and Plonsky2016; Myers, Reference Myers1990), and visual inspection of residual plots did not reveal any obvious deviations from homoscedasticity or normality.

Results

Figure 1 and Figure 2 show boxplots of reaction times from the phrasal acceptability judgment task in milliseconds for each member of the pairs in the high and low cutoff bins, respectively. Generally, participants in both groups were faster when judging the high-frequency phrase in each pair, and the NSs were faster than the NNSs.

Figure 1. Boxplots for reaction times (ms) for each stimuli pair in the high cutoff bin. The order of the pairs on the x-axis is based on the frequency of the high-frequency member of each pair; the high-frequency member of the pair on the far left has the lowest frequency among all the high-frequency members of all pairs. NS = native English speaker. NNS = non-native English speaker.

Figure 2. Boxplots for reaction times (ms) for each stimuli pair in the low cutoff bin. The order of the pairs on the x-axis is based on the frequency of the high-frequency member of each pair; the high-frequency member of the pair on the far left has the lowest frequency among all the high-frequency members of all pairs. NS = native English speaker. NNS = non-native English speaker.

Table 2 shows the average reaction times from the two participant groups. The last column indicates the difference in reaction time between the high-frequency condition and the low-frequency condition (the baseline frequency category) in each cutoff bin in each participant group. In the high cutoff bin, in which the cutoff point for classifying a phrase as a high-frequency phrase or a low-frequency phrase was 12.00 occurrences per million, the native speaker mean reaction time for high-frequency phrases was about 84 ms shorter than the mean reaction time for low-frequency phrases. This difference was greater than the difference of 60 ms in Arnon and Snider’s (Reference Arnon and Snider2010) study, in which native speaker mean reaction time for high- and low-frequency phrases were 1040 ms and 1100 ms, respectively. In the low cutoff bin, in which the cutoff point was 1.00 occurrence per million, the mean reaction time for high-frequency phrases was approximately 55 ms shorter than the mean reaction time for low-frequency phrases in the native speaker group. This was smaller than the 66-ms difference in Arnon and Snider’s (Reference Arnon and Snider2010) study, in which respective mean reaction time for high- and low-frequency phrases were 1040 ms and 1100 ms. Moreover, while the NNSs were generally slower than the NSs, these NNSs also reacted faster to high-frequency phrases than to low-frequency phrases. In the high and the low cutoff bins, respectively, the mean reaction times for high-frequency phrases were about 116 ms and 94 ms shorter than the mean reaction times for low-frequency phrases.

Table 2. Average reaction times in milliseconds from the phrasal acceptability judgment task (SD in parentheses)

Note: NS = native English speaker. NNS = non-native English speaker.

Table 3 reports results from the regressions for the high cutoff bin. There were significant main effects of phrase frequency, χ² (1) = 7.13, p = .008, participant group, χ² (1) = 61.65, p < .001, block order, χ² (1) = 115.45, p < .001, and the number of characters in the target phrases, χ² (1) = 21.74, p < .001. The regression coefficient (β) for each predictor indicates the change in reaction times on the base-10 logarithmic scale as a result of a 1-unit change in the predictor. In the case of binary predictors, the exponential value (10^β) expresses the average multiplicative change in reaction time between the nonreference category (e.g., high) and the reference category (e.g., low). Similarly, for continuous predictors, which were standardized, the exponential value expresses the average multiplicative change in reaction time associated with a 1 SD change in the predictor.

Table 3. Mixed-effects model results for the high cutoff bin in the phrasal acceptability task

Note: R ² marginal = .26. R ² conditional = .54. SE = standard error. NS = native English speaker.

As Table 3 shows, both groups demonstrated sensitivity to phrase frequency. On average, reaction times to high-frequency phrases were 0.91 times the reaction times to low-frequency phrases. That is, participants were about 9% faster when judging the acceptability of high-frequency phrases. Moreover, the NNSs were on average 33% slower than the NSs, and participants were on average 9% faster when they did the second experiment block, whether it was Block A or Block B, possibly due to greater task familiarity. While the counterbalanced design equally distributed the familiarity effects between the two blocks, with block order as a control variable, the portion of variance in reaction time due to task familiarity was accounted for statistically.

In addition, a 1 SD increase in the number of characters in the target phrases corresponded to about a 10% increase in reaction times, indicating that the participants had to spend more time reading the phrases. The interaction between phrase frequency and group, χ² (1) = 0.04, p = .842, and the interaction between phrase frequency and block order, χ² (1) = 0.56, p = .453, were not significant, and neither were the frequencies of the last word, χ² (1) = 0.04, p = .842, the last two words, χ ² (1) = 0.59, p = .444, and the last three words, χ² (1) = 1.22, p = .269. As in previous similar research (Gyllstad & Wolter, Reference Gyllstad and Wolter2016), the MUMIn function in R was used to obtain R ² values. This function provides two R ² values: marginal and conditional. The former is associated with the fixed effects, listed in this table, and the latter reflects the fixed and the random effects combined. In this model, the fixed effects and the random effects together explain about 54% of the variance in the participants’ reaction times.

As Table 4 shows, a similar result pattern was obtained from the low cutoff bin. The effects of phrase frequency, χ² (1) = 4.78, p = .029, participant group, χ² (1) = 85.79, p < .001, block order, χ² (1) = 125.92, p < .001, and the number of characters in the target phrases, χ² (1) = 19.38, p < .001, were significant. Therefore, both groups were sensitive to phrase frequency. On average, they were 10% faster when judging high-frequency phrases, and the NNSs were approximately 42% slower than the NSs. Reaction times in the second experiment block were on average 12% shorter than in the first block, and a 1 SD increase in the number of characters in the target phrases corresponded to an approximately 9% increase in reaction times. The interaction between phrase frequency and group, χ² (1) = 0.15, p = .696, and the interaction between phrase frequency and block order, χ² (1) = 0.39, p = .531, were not significant. Moreover, frequencies of the last two words, χ² (1) = 1.32, p = .250, and the last three words, χ² (1) = 0.59, p = .442, did not reach significance, while frequencies of the last words, χ² (1) = 3.75, p = .053, was close to being significant.

Table 4. Mixed-effects model results for the low cutoff bin in the phrasal acceptability task

Note: R ² marginal = .31. R ² conditional = .56. SE = standard error. NS = native English speaker.

Experiment 2

This experiment tested frequency effects on production of four-word compositional phrases. An elicited oral production task was used because this task demonstrated phrase frequency effects in two previous studies (Arnon & Priva, Reference Arnon and Priva2013; Bannard & Matthew, Reference Bannard and Matthews2008) and helped increase the comparability between the two experiments in the current study (i.e., elicitation of the same target phrases). Moreover, as in these previous studies, language production was operationalized as the phonetic durations of the first three words in the target phrases, with the assumption that the same words should be produced faster in a more frequent phrase (e.g., don’t have to worry) than in a less frequent phrase (e.g., don’t have to wait).

Results

Figure 3 and Figure 4 show boxplots of production durations of the target segments in milliseconds for each target pair in the high and the low cutoff bin, respectively. Overall, the target segment in the high-frequency member of each pair was produced faster than the target segment in the low-frequency member, and the NSs were faster than the NNSs.

Figure 3. Boxplots for production durations (ms) of the target segments for each stimuli pair in the high cutoff bin. The order of the pairs on the x-axis is based on the frequency of the high-frequency member of each pair; the high-frequency member of the pair on the far left has the lowest frequency among all the high-frequency members of all pairs. NS = native English speaker. NNS = non-native English speaker.

Figure 4. Boxplots for production durations (ms) of the target segments for each stimuli pair in the low cutoff bin. The order of the pairs on the x-axis is based on the frequency of the high-frequency member of each pair; the high-frequency member of the pair on the far left has the lowest frequency among all the high-frequency members of all pairs. NS = native English speaker. NNS = non-native English speaker.

Table 5 shows the average production duration for the target segments in each condition. In the NS group, in both high and low cutoff bins, production durations for the target segments inside high-frequency phrases were shorter than production durations for the target segments inside low-frequency phrases. That is, in the two respective bins, the mean production durations for the target segments in high-frequency phrases were about 18 ms and 14 ms shorter than the mean production durations for the target segments in low-frequency phrases. Similarly, in the NNS group, the mean production durations for the target segments in high-frequency phrases were approximately 18 ms and 17 ms shorter than the mean production durations for target segments in low-frequency phrases, respectively.

Table 5. Average production durations of the target segments in milliseconds from the elicited production task (SD in parentheses)

Note: NS = native English speaker. NNS = non-native English speaker.

However, as shown in Table 6, results from the regression models for the high cutoff bin suggested that phrase frequency did not significantly predict production durations of the target segments. Although the coefficient for phrase frequency was negative as expected, phrase frequency effects did not reach significance, χ² (1) = 2.35, p = .125. There were, however, significant main effects of participant group, χ² (1) = 27.93, p < .001, block order, χ² (1) = 16.76, p < .001, and the number of syllables in the target segments, χ² (1) = 12.77, p = < .001. That is, the NNSs were on average 14% slower than the NSs. Participants were on average 2% faster when they did the second experiment block possibly due to greater task familiarity. In addition, a 1 SD increase in the number of syllables in the target segments led to about a 5% increase in production durations. The interaction between phrase frequency and group, χ² (1) = 0.47, p = .493, and the interaction between phrase frequency and block order, χ² (1) = 0.62, p = .432, were not significant. The frequencies of the last word, χ² (1) = 0.28, p = .595, and the last three words, χ² (1) = 1.22, p = .270, also did not reach significance. However, the effects of frequency of the last two words in the target segment were significant, χ² (1) = 6.60, p = .010. The negative coefficient for this predictor suggested that a 1 SD increase in the log frequency of the last two words in the target phrases (e.g., to worry in don’t have to worry) corresponded to a 5% decrease in the production durations of the target segment (e.g., don’t have to). In light of usage-based approaches (e.g., Bybee, Reference Bybee2010), this could mean that the higher frequency, and thus the stronger relation, between the last word in a target segment (e.g., to) and the last word in a phrase (e.g., worry) made the participants produce the last word in the target segment faster, leading to shorter production durations of the target segment.

Table 6. Mixed-effects model results for the high cutoff bin in the elicited production task

Note: R ² marginal = .22. R ² conditional = .70. SE = standard error. NS = native English speaker.

Table 7 shows results for the low cutoff bin. Similarly, the coefficient for phrase frequency was negative, but phrase frequency effects did not reach significance, χ² (1) = 3.21, p = .073. There were significant main effects of participant group, χ² (1) = 35.96, p < .001, block order, χ² (1) = 12.44, p < .001, and the number of syllables in the target segments, χ² (1) = 122.84, p < .001. That is, the NNSs were on average 17% slower that the NSs. Moreover, participants were on average 1% faster when they did the second experiment block, and a 1 SD increase in the number of syllables in the target segments led to about a 15% increase in production durations. The interaction between phrase frequency and block order was significant, χ² (1) = 3.98, p = .046. The negative coefficient indicates that, in the second block, participants produced the target segments in high-frequency phrases 2% faster than the target segments in low-frequency phrases. The interaction between phrase frequency and group, χ² (1) = 0.39, p = .527, was not significant, and neither were the frequencies of the last two words, χ² (1) = 0.23, p = .636, and the last three words, χ² (1) = 0.97, p = .325.

Table 7. Mixed-effects model results for the low cutoff bin in the elicited production task

Note: R ² marginal = .45. R ² conditional = .81. Log frequency of last word in the target phrase was excluded to reduce collinearity among predictors. SE = standard error. NS = native English speaker.

General discussion

As Arnon (Reference Arnon2015) pointed out, “[f]requency effects are not interesting in and of themselves. They are interesting because they reveal something about the [language] learning mechanisms …” (p. 274). The current study investigated whether frequency effects can be observed in the processing of four-word English compositional processing in adult NSs and NNSs. Such effects are predicted in usage-based approaches, which attribute L1 and L2 acquisition to the interaction between domain-general human cognitive processes and input and predict that frequency affects processing of linguistic units at all levels (Bybee, Reference Bybee2010; Goldberg, Reference Goldberg2006; Tomasello, Reference Tomasello and Bavin2009). In terms of receptive processing, the frequency effects observed from the NSs in Experiment 1, as well as from NSs in previous studies on four-word phrases (Arnon & Snider, Reference Arnon and Snider2010; Hernández et al., Reference Hernández, Costa and Arnon2016), morphemes (Ambridge et al., Reference Ambridge, Kidd, Rowland and Theakston2015), single words (e.g., Diependaele et al., Reference Diependaele, Lemhöfer and Brysbaert2012), idioms (Nippold & Rudzinski, Reference Nippold and Rudzinski1993), and two-word phrases (e.g., Gyllstad & Wolter, Reference Gyllstad and Wolter2016), therefore provide empirical evidence supporting usage-based researchers’ claim. Regarding NNSs, the results from Experiment 1 and recent findings from Hernández et al. (Reference Hernández, Costa and Arnon2016) extend empirical support for frequency effects on English L2 receptive processing of single words (e.g., Diependaele et al., Reference Diependaele, Lemhöfer and Brysbaert2013; Whitford & Titone, Reference Whitford and Titone2012) and two-word collocations (e.g., Gyllstad & Wolter, Reference Gyllstad and Wolter2016; Sonbul, Reference Sonbul2015; Wolter & Gyllstad, Reference Wolter and Gyllstad2013) to longer compositional sequences. These likewise appear to corroborate the proposal in usage-based approaches to L2 acquisition (Ellis, Reference Ellis, Gass and Mackey2011, Reference Ellis, Trousdale and Hoffmann2013).

Given these findings, the next question is how to develop a corresponding L1 and L2 representation and processing psycholinguistic model. The results from Experiment 1 are compatible with usage-based models of language acquisition in which frequency of input affects language learning and construction representation at all levels (Bybee, Reference Bybee2010; Goldberg, Reference Goldberg2006; Tomasello, Reference Tomasello and Bavin2009). As Arnon and Snider (Reference Arnon and Snider2010) noted, an adequate model has to address several key issues, such as how an encounter of a specific lexical sequence is counted as an instance of multiple more abstract sequences, and the relationship between the representation of multiword phrases (e.g., don’t have to worry), the subparts (e.g., to worry), and the more abstract linguistic units related to the subparts (e.g., an infinitive clause). Moreover, according to Ibbotson (Reference Ibbotson2013), such a model needs to accommodate cognitive processes such as chunking and categorization, and should be able to expand and organize itself. Undoubtedly, a model for L2 speakers will be even more complicated due to various additional factors (e.g., L1 influence and L2 proficiency). Therefore, similar to the results from Hernández et al. (Reference Hernández, Costa and Arnon2016), the findings in Experiment 1 are in line with a unified connectionist model of L1 and L2 learning (MacWhinney, Reference MacWhinney, Robinson and Ellis2008), in which the processes underlying L1 and L2 acquisition share similarities and are not separable. In such models, both L1 and L2 acquisition rely on common cognitive processes, such as memory retention of word co-occurrences and chunking, and statistical information in input, including frequency, affects processing of L1 and L2 linguistics units at all levels, including compositional phrases. Moreover, there does not need to be a qualitative difference between high- and low-frequency phrases; the entrenchment and processing difference is quantitative, resulting from different frequencies of previous phrase encounters.

Regarding phrase production, Experiment 2 in the current study lent weak support for phrase frequency effects from the adult NSs and NNSs, who demonstrated such effects in the receptive task in Experiment 1. To date, results from the relevant L1 research based on elicited phrase production durations, including the current study, have been mixed, and besides the current study, Ellis et al.’s (Reference Ellis, Simpson-Vlach and Maynard2008) study seems to be the only existing related study conducted with NNSs. As discussed, cross-study methodological differences could have contributed to the incongruent findings, and such differences should be investigated in future research, particularly given the relatively limited research on frequency effects on phrase production durations, especially in NNSs.

Perhaps it should also be pointed out that, unlike elicited production research, research based on four-word phrase production in NSs’ spontaneous speech has more consistently reported frequency effects (Arnon & Priva, Reference Arnon and Priva2013, Reference Arnon and Priva2014). This consistency, together with frequency effects based on L1 spontaneous single-word production (e.g., Bell et al., Reference Bell, Jurafsky, Fosler-Lussier, Girand, Gregory and Gildea2003; Bybee & Scheibman, Reference Bybee and Scheibman1999), might suggest that the type of speech investigated (elicited vs. spontaneous) is another methodological factor contributing to whether phrase frequency effects can be observed in empirical research. If frequency effects in phrase production are related to “activation of multi-word lemmas” (Arnon & Priva, Reference Arnon and Priva2013, p. 366), in the current study it might be possible that by the time each target phrase disappeared from the computer screen in Experiment 2, the multiword lemmas had already been activated, so the difference in production durations between high- and low-frequency phrases was reduced. Possibly, compositional phrase frequency effects will be observed more consistently if spontaneous speech from NSs and NNSs is the subject of investigation. As in the case of comprehension, if there is more evidence for frequency effects on phrase production, such sensitivity will entail the need for a development of a psycholinguistic model that accommodates both word and multiword phrase frequency in speech production. As Arnon and Priva (Reference Arnon and Priva2013) suggested, such a model can be an expanded version of the connectionist models of L1 production (e.g., Chang, Reference Chang2002; Chang, Dell, & Bock, Reference Chang, Dell and Bock2006). An adequate model must accommodate activation and competition among single words and multiword phrases during phrase production.