Input frequency and construction interference interactions in L2 development

Abstract

The interaction between input frequency and constructional interference receives little attention in second language (L2) research. Two studies were conducted to test the effect of this interaction. Study 1 examined effects of both Zipfian frequency (ZF) and balanced frequency (BF) on L2 learning of English subject-extracted relative clauses (SRs). Participants learned SRs and finished a picture description task at pretest, posttest and delayed posttest. Results suggest that ZF was not superior to BF for learners to use SRs. Study 2 tested effects of ZF and BF on L2 learning of English object-extracted relative clauses (ORs). The learning and testing procedures were the same as those in Study 1, but different from Study 1 participants either with or without previous exposure to SRs learned ORs. Results show that ZF was not superior to BF for learners without previous exposure to SRs to use ORs, but more beneficial than BF for those with previous exposure to SRs to use ORs to describe pictures. These findings demonstrate that ZF was more effective than BF in inhibiting SR interference.

Keywords

balanced frequency interference object-extracted relative clauses second language development subject-extracted relative clauses Zipfian frequency

I Introduction

Usage-based linguistics assumes that language learning is a statistically-driven process, during which learners’ linguistic knowledge is continuously sensitive and experientially adaptive to the structure of input, and that as input accrues, learners are able to analyse the commonality of all instances, thereby gradually abstracting a construction pattern (Bybee, 2008; Ellis, 2002; Pajak et al., 2016). The formation of a linguistic pattern is mainly dependent on language input that learners receive. The structure of input, however, often consists of various factors that may deferentially influence the developmental trajectory of second language (L2) acquisition. Input frequency and construction interference are important variables that have been extensively researched. For the input frequency, Zipfian and balanced frequencies, two special input distributions, have recently been under close scrutiny. In the Zipfian distribution, a typical instance possesses a lion’s share in the occurrence of all instances, while in the balanced distribution, all instances occur approximately with equal possibility. The two distributions have been theorized to play differing roles at distinct stages of language development (Ellis et al., 2013; Ellis and Cadierno, 2009; Goldberg, 2006; Goldberg et al., 2004; Year and Gordon, 2009). Construction interference, as evinced by the fact that learners transfer related linguistic knowledge to the process where the target construction is learned, is argued to be able to either hinder or facilitate L2 learning (Larsen-Freeman, 2013). Hindrance from other constructions often manifests itself as intralingual errors: the confusion L2 learners experience when confronting patterns within structures of the target language (Scovel, 2001). Errors of this kind in L2 learning of morphological morphemes, synonyms, verb–argument structures, and relative clauses may result from learners’ false analogy, overgeneralization (Schachter and Celce-Murcia, 1977), incomplete rule application (Richards, 1971), blocking from alternative constructions (Ellis and Sagarra, 2011), and inability to distinguish similar constructions (McDonough and Trofimovich, 2013). Facilitative roles of related constructions in learning the target constructions, albeit less extensively researched, have been detected in Abbot-Smith and Behrens (2006). For instance, the learning of German copula sein and werden-passive was supported by the auxiliary sein and the copula werden.

Despite extensive separate research into the effects of these factors, it remains largely unexplored whether and how they interact with one another to influence L2 development. This issue is theoretically significant since one of the primary purposes of L2 learning studies is to extensively investigate how potential factors and their interactions impact the way the target language is learned. The current study is thus motivated to explore how these variables conspire to influence L2 acquisition of English relative clauses.

II Frequency distribution in language learning: Zipfian vs. balanced

Zipfian law, which states that the frequency of a word is approximately inversely proportional to its rank, has been found linguistically universal, such as in words, phrases, and abstract constructions (Ellis et al., 2013; Zipf, 1935). This law has been theorized to facilitate language development due to its distributional property, i.e. a typical instance that possesses a lion’s share of the occurrence of all instances. The recurrence of this instance not only serves as an anchor for the acquisition of other less frequent instances (Ellis and Collins, 2009; Ellis et al., 2013), but consistently provides low-variance input initially as well, making the construction more tangible (Crossley et al., 2013; Ellis, 2006a; Ellis and Ferreira-Junior, 2009; Goldberg et al., 2004). In contrast, balanced input contains exemplars that have equal token frequency. Theoretically, learners exposed to balanced input are less likely to rely on the anchor as afforded by the most frequent instance in the Zipfian input, therefore having difficulty establishing a linguistic category at the initial stage. This conjecture has gained support from Goldberg et al. (2004) and Ellis and Ferreira-Junior (2009), who analysed English children’s and L2 learners’ use of verb–argument constructions (VACs), finding that English children and L2 learners consistently used particular verbs in these VACs. These findings were taken as the role of Zipfian frequency: learners repeatedly encounter typical instances, and thereupon are able to employ this instance to build up a linguistic category (Goldberg, 2006). The advantage of Zipfian input over balanced input has also been found in research on artificial language learning. In Goldberg et al. (2004, 2007), English children and adults were exposed to the novel NP1-NP2-VP structure. Zipfian and balanced input were distinguished for children. In the Zipfian input, one verb occurred in 50% of all instances, and other different verbs occurred in the remaining 50%, while in the balanced input, different verbs occurred evenly in all instances. Results showed that Zipfian input was superior to balanced input when children learned the target construction. Zipfian first and Zipfian random input were tested for adults. In the former, adults were exposed to one and the same verb 8 times in succession and twice to four other verbs; in the latter, adults received 8 times of a verb randomly interspersed with the other eight verbs. Results indicated that Zipfian first input was superior to Zipfian random input for adults to generalize the target construction. Kidd et al. (2006) found that children tended to resort to the high-frequency verb when learning the sentential complement construction. Wulff et al. (2009) also detected that a verb frequently used in the tense-aspect construction could improve tense-aspect learning due to the recurrence of this verb.

The advantage of Zipfian input over balanced input, however, have not been found in most of the studies regarding L2 learning of VACs. Year and Gordon (2009) explored Korean English-L2 learners’ acquisition of English ditransitives under these two conditions, finding that the learners receiving balanced input outperformed those receiving Zipfian input at the delayed posttest. Nakamura (2012) utilized Casenhiser and Goldberg’s (2005) paradigm to investigate the Japanese speakers’ learning of the NP1-NP2-VP construction and ergative construction. Results indicated that learners receiving balanced input did better than those receiving Zipfian input in producing the NP1-NP2-VP construction, while learners receiving the two types of input did not differ in comprehending these two constructions. McDonough and Nekrasova-Becker (2014) also found that balanced input was superior to Zipfian first and Zipfian random input during the process where Thai speakers learned English ditransitives, while non-significant difference was observed between Zipfian first and Zipfian random input. Further, McDonough and Trofimovich (2013) examined how input frequency interacted with the learning mode (rules given in the deductive condition vs. rules not given the inductive condition) in the process where Thai university students detected the Esperanto transitive pattern, with results that only the learners receiving balanced input under the deductive condition detected the pattern of Esperanto transitives. Fulga and McDonough (2016) tested whether L1 background and presentation visuals (color vs. black-and-white) interacted with input frequency in the development of Esperanto transitives. Significant effect was observed for L1 background instead of visuals and Zipfian/balanced frequency in the L2 learners’ detection of Esperanto transitives. However, Zhang and Dong (2016) investigated how Zipfian and balanced input influenced Chinese speakers’ acquisition of English prenominal adjective participles, finding that Zipfian input was more effective than balanced input for learners to reject the ungrammatical prenominal adjective participles. Dong and Zhang (2017) probed how Zipfian and balanced input affected L2 development of English pronominal adjective participles and object-extracted relative clauses. They observed that Zipfian input was superior to balanced input in L2 learning of English pronominal adjective participles, confirming the conclusion of Zhang and Dong (2016), while the two distributions played no differential role in L2 development of object-extracted relative clauses. They concluded that the two distributions may not differ in the L2 learners’ development of abstract constructions.

These results demonstrate that Zipfian input may be no more effective than balanced input in L2 development of VACs and other abstract constructions. It has been argued that the lack of significant effect for Zipfian input on L2 learning may be attributable to learning environment (McDonough and Nekrasova-Becker, 2014) and learners’ reliance on explicit reasoning (Fulga and McDonough, 2016; Nakamura, 2012; Year and Gordon, 2009). McDonough and Trofimovich (2013) assumed that Zipfian input may not play a facilitative role when L2 learners rely upon explicit reasoning as opposed to implicit L1 learning based mainly on input as in Goldberg et al. (2004, 2007). This assumption needs to be tested when L2 learners often receiving explicit teaching learn other constructions. Other factors that may help explain inconsistent findings in L1 and L2 learning have to do with methodological differences, target constructions, learning and testing requirements. For the methodological differences, both type and token frequencies in these studies differed substantially. The token frequency of each verb in the balanced input of Year and Gordon (2009) was eight, almost four times as high as that in Goldberg et al. (2004), Zhang and Dong (2016), and Dong and Zhang (2017). Also, the token frequency of exemplars in the Zipfian input varied, with eight times in Goldberg et al. (2004) and McDonough and Nekrasova-Becker (2014), twenty-four times in Year and Gordon (2009), seven times in Zhang and Dong (2016), twelve times in McDonough and Trofimovich (2013), and fifteen times in Fulga and McDonough (2016). Great variance in token frequencies in the intended input is likely to influence the observed effects of input distributions, since Zhang and Ma (2014) found that six occurrence events are enough for L2 learners to acquire a new instance of a VAC. Investigations into the effects of Zipfian and balanced frequencies should take this into consideration. It is important to explore what effects can be observed when the skewed frequency of an exemplar in Zipfian input is approximately six times and the token frequency of an exemplar in balanced input is less than six times. For the target constructions, NP1-NP2-VP structures (Goldberg et al., 2004, 2007; Nakamura, 2012), Esperanto transitives (Fulga and McDonough, 2016; McDonough and Trofimovich, 2013), sentential complements (Kidd et al., 2006), tense-aspect constructions (Wulff et al., 2009), ditransitives (McDonough and Nekrasova-Becker, 2014; Year and Gordon, 2009), and prenominal adjectival participles (Zhang and Dong, 2016) were involved. Although the focus of these studies was on statistical learning from input, the use of real or artificial language may lead to different results. Considering that artificial languages are more grammar-based and less semantically- or pragmatically-constrained than real languages, so more investigations into other different constructions are warranted to comprehensively evaluate the effectiveness of the two input distributions in L2 development. For learning and testing requirements, McDonough and Trofimovich (2013) only explored the learners’ ability to identify the object nouns of the Esperanto transitives instead of their knowledge of these transitives. Learners in the deductive group were a priori informed that the suffix -n marked the accusative case of the Esperanto transitives, and consequently they had an advantage over those in the inductive group. On the other hand, the group receiving Zipfian input may have had difficulty detecting the object nouns because morpheme occurrence is easier to detect from balanced input (Krajewski et al., 2011). Additionally, the inability of learners receiving skewed and balanced input in the inductive group was probably caused by the complexity of the Esperanto transitives. That is, both SVO and OVS may have overloaded the participants’ short-term memory. In this sense, further attention should be devoted to exploring how the two distributions help L2 learners learn the target construction in an inductive way.

Based on these observations, the first purpose of the present study is to expand the existing L2 research via comparing the effectiveness of Zipfian and balanced input in L2 learning of other constructions, say English relative clauses. More importantly, L2 learners always encounter constructions similar in form or function or both in daily use. Similar items often incur information interference in the formation of a category. Such information interference may also apply to L2 development in that the Associative-Cognitive CREED model predicts that the process of L2 acquisition is no different from that of the learning of any other type of information (Ellis, 2006b). More attention needs to be devoted to determining the way in which learners manage to untangle such interference under different input conditions. Studies in this line may give insights into instructed L2 learning, and is thus of vital significance.

III Construction interference

Construction interference resulted from false analogy, misanalysis, overgeneralization (Schachter and Celce-Murcia, 1977), incomplete rule application (Richards, 1971), blocking from competing constructions (Ellis and Sagarra, 2011), and inability to distinguish different constructions when presented simultaneously has been extensively documented in the L2 literature. In this part, we reviewed five studies that are directly related to our focus. Goldberg and Casenhiser (2008) reported an unpublished study, which tested the effectiveness of skewed and balanced input in the children’s learning of novel VACs. Children learned new instances of the transitive construction interspersed among the novel VACs under both skewed and balanced conditions. The children were familiar with the transitive construction in advance. They found that the children exposed to both skewed and balanced input performed significantly above chance in assigning transitive meaning to the transitive construction at test, while those under either of the condition failed to recognize the form-meaning mapping of the novel VACs. The authors attributed the null effects of skewed/balanced input to interference of the transitive construction. In a similar study, McDonough and Trofimovich (2013) conformed Goldberg and Casenhiser’s (2008) conclusion when Thai speakers learned the Esperanto transitives, characterized by the accusative suffix and different word order (SVO, OVS). During learning, participants learned 24 instances under skewed or balanced input following either inductive (without rule) or deductive (with rule) instructions. During testing, they heard 20 sentences (10 SVOs, 10 OVSs) with new nouns and then identified the object. Results suggested that only the group receiving balanced input with deductive instructions found the pattern. The authors also attributed L2 learners’ failure in pattern-finding to the influence of the SVO structure, with which Thai university students were familiar. One possibility for such null effects in Goldberg and Casenhiser (2008), and in McDonough and Trofimovich (2013) is that the target construction and the familiar construction were learned simultaneously. This may make the learning task more challenging than learners could handle, as reflected in their floor-effect performance at test. It remains less clear about whether and how these input distributions interact with different constructions when the learning task was not that difficult, for instance, different constructions presented separately instead of simultaneously. What is more, unlike Goldberg and Casenhiser (2008), and McDonough and Trofimovich (2013), language instructors in classroom settings seldom mix up different constructions but present them sequentially in the same or different learning sessions. In this sense, it is pedagogically significant to probe the effectiveness of Zipfian and balanced input at facilitating the learning of similar constructions that are presented sequentially.

Morris et al. (2000) proposed a connectionist model for wh-question learning, finding the ‘construction-conspiracy’ effect: the model can be exclusively generalized to untrained wh-instances when the model learned syntactically-/semantically-related structures. Both positive and negative influence were observed by Abbot-Smith and Behrens (2006), who analysed utterances produced by a German boy aged from 2 to 5. They investigated L1 acquisition of German passive and future constructions, both of which contained a verb with either the auxiliary sein (‘to be’) or werden (‘to become’). Results demonstrate that the boy acquired the sein- before the werden-passive, that the auxiliary sein was supported by his prior acquisition of the copula sein, whereas the werden-passive supported one werden copula construction, and that the boy acquired the werden-future very slowly because of interference from a semantically identical construction. The positive influence from the copula sein or the werden-passive when the auxiliary sein or the copula werden was learned has confirmed the ‘construction-conspiracy’ effect. The negative influence from the related construction in learning the werden-future may be explained by the preemption principle (Clark and Clark, 1979), which predicts that learners tend to consider two constructions to be identical in meaning initially if they occur in similar contexts. When learners are capable of differentiating the meaning between these constructions, they will use them properly. This principle has been born out L1 learning studies (e.g. Ambridge et al, 2015; Tomasello, 2003). It seems that Morris et al.’s (2000) and Clark and Clark’s (1979) predictions on construction interference are different. According to Morris et al. (2000), the acquisition of a construction will be facilitated by the construction when the two constructions share important elements or similar functions. However, Clark and Clark (1979) have predicted that the learning of a construction will be hindered by another one when the two constructions express similar meanings or their meanings are indistinguishable for learners initially. These two assumptions have been widely assumed rather than treated as a central topic in L2 research, so another aim of the present study is to examine whether, and to what extent, different input conditions interact with different constructions to impact the result of L2 learning.

IV Target constructions

Both English subject-extracted relative clauses (SRs) and object-extracted relative clauses (ORs) were targeted based on the following reasons. First, English belongs to right-branching languages, while Chinese belongs to left-branching languages. The difference regarding SRs and ORs in the two languages are illustrated in Examples (1) and (2).

(1) Subject-extracted relative clauses

The woman [that_ is eating the apple] is his mother.

[那个吃苹果的]女人是他的妈妈。

[chī píng guǒ de] nǚ rén shì tā de mā mā

[Eating the apple] woman is his mother

(2) Object-extracted relative clauses

The apple [that the woman is eating _] is red.

[那个女人吃的]苹果是红色的

[nà gè nǚ rén chī de] píng guǒ shì hóng sè de

[The woman is eating] apple is red

In Example (1) the English SR clause [that_ is eating the apple] modifies its antecedent woman postnominally, while the corresponding Chinese relative clause [那个吃苹果的] (‘eating the apple’) modifies its head 女人 (‘woman’) prenominally. This is also true for the English OR clause [that the woman is eating _] and its corresponding Chinese relative clause [那个女人吃的] (‘The woman is eating’) in Example (2). From the cross-linguistic difference regarding these two relative clauses, we postulate that during the formation of English relative clauses, Chinese speakers are supposed to find the pattern for both clauses mainly based on input instead of making use of L1 knowledge. Such differences between Chinese and English have been found to spell acquisition difficulty for Chinese English-L2 learners (Cai and Wu, 2006). Second, the Accessibility Hierarchy suggests an implicational relationship between SRs and ORs (Gass 1979; Keenan and Comrie 1977), which has potential implications in language acquisition, as Eckman’s (1984) typological markedness hypothesis has suggested that acquisition ease should follow SRs > ORs, with less accessible or lower positions being ‘more marked’ and more difficult to acquire than more accessible or higher positions in the Accessibility Hierarchy (AH). Eckman’s assumption has been borne out by Gass (1979), who tested L2 learners of different L1 backgrounds, finding that L2 learners’ accuracy rates of using English relative clauses were consistent with the AH. Such a relation may help us observe the potential interference of SRs on the learning of ORs. Third, according to the teaching syllabus of English for Chinese students, SRs are learned before ORs, so testing the interference of SRs on ORs could give more insight into classroom-based language teaching. Our findings are expected to reflect the possible learning outcome triggered by the instructional intervention. In light of the assumed effect of Zipfian input, repeated high token frequencies constitute low-variance input and therefore help learners release more cognitive resources to cope with different constructions that are similar in from or function than does balanced input. It is likely that Zipfian input might be more effective than balanced input to help learners learn ORs after they learned SRs.

V Research questions

Based on the literature and analysis of English SRs and ORs, we conducted two studies to answer the following three research questions.

Does the effectiveness of Zipfian input and balanced input at promoting L2 learning of English SRs differ?

Does the effectiveness of Zipfian input and balanced input at promoting L2 learning of English ORs differ?

Does the effectiveness of Zipfian input and balanced input at promoting L2 learning of English ORs differ after learners learned SRs?

VI Study 1

Study 1 was intended to answer the first research question.

1 Method

a Participants

Sixty third-year middle school students were recruited. They came from rural areas of northwest China with almost no access to other English media except course books. They were randomly assigned to three groups, with 20 students in each, labeled as the Zipfian frequency (ZF) group, balanced frequency (BF) group, and control group (CG). They had received formal English teaching for 2 years in primary school and 2.5 years in middle school. Each weekday, they received 45-minutes explicit English teaching. None of them had been to English-speaking countries. According to their instructor, they had not learned SRs and ORs before. A search by the two authors for the usage of relative clauses in the participants’ textbooks found only one instance, the way you speak. It can be assumed that participants had little knowledge of SRs and ORs. This assumption was further tested through an acceptability judgment task and a picture description task administered to another 10 students from the same population. The acceptability judgment task consisted of 10 SRs (5 grammatical and 5 ungrammatical) and 10 ORs (5 grammatical and 5 ungrammatical) on a 7-point scale, with 1 = totally unacceptable, 4 = I’m not sure, and 7 = completely acceptable. Results suggest that they gave the score of 4 to 91% items, 1 to 1% items, 3 to 6% items, and 7 to 5% items. The picture description task consisted of 6 pictures for SRs and 6 for ORs (the same as those used in Study 1 and Study 2) to test the students’ knowledge of SRs and ORs, but they did not produce any instance of English SRs or ORs. All participants had learned the five English basic sentence structures (i.e. S+V, S+ Linking-V, S+V+O, S+V+O+O, S+V+O+C),¹ so they were developmentally ready for learning English relative clauses. Results of the one-way ANOVA on the scores of all participants’ recent midterm test suggested they were at the same English proficiency level (F(2, 57) = 1.637, p = .204). Their biodata is presented in Table 1.

Table 1.

Biodata of the participants.

Groups	n	Age	Male; Female	Midterm test scores
Groups	n	Age	Male; Female	Mean	SD
ZF	20	14–16	13; 7	81.40	4.97
BF	20	13–15	11; 9	82.35	3.03
CG	20	13–15	8; 12	79.50	6.58

Notes. ZF = Zipfian input; BF = balanced input; CG = control group.

b Data collection

The investigation was undertaken in regular classes. On the first day, all participants finished the picture description task at the pretest. On the second day, the experimental groups learned 6 picture-aided English SRs, while the CG attended regular classes. After learning, participants immediately took the posttest. In order to examine whether there was a long-term learning effect triggered by the treatment, the delayed posttest was carried out one week after the posttest. At each testing session, the picture description task was used to evaluate the extent of English SR acquisition. To minimize the possibility of practice effect, the first version of the task was used at pretest and delayed posttest, while the second version was used at the posttest. Participants spent approximately 20 minutes finishing each test.

c Training

The two experimental groups learned the SR with matched pictures that expressed the meaning of each construction via images. The SR and matched pictures were presented on slides. Each picture described a scene where someone was doing something. The construction learning consisted of two phases (see Figure 1). On Phase 1, Picture A together with the two sentences (e.g. the dog is chasing a man and the dog is black) that described this picture appeared on the side. Participants were informed that the sentences described this picture. They were then asked to look at the picture and read the two sentences once after the second author. On Phase 2, Picture B and a SR that was made up of the two sentences describing Picture B (the dog that is chasing the man is black) were presented. In Picture B, an arrow was provided to ensure that the participants’ attention would be directed to the entity that would be modified by the SR. No explicit explanations were provided as to the formation of SRs. During learning, participants receiving Zipfian input were first exposed to a sentence containing a SR 7 times, since Zhang and Ma (2014) found that 6 occurrences or above can guarantee learners’ acquisition of a new construction. Then, they learned other five sentences that contained five different SRs, with each presented once. Participants receiving balanced input were exposed to six sentences that consisted of 6 different SRs, with each presented twice. The learning session lasted about 20 minutes. The frequency and presentation order of all instances for the two groups were 7-1-1-1-1-1 and 2-2-2-2-2-2.

Figure 1.

An example of a subject-extracted relative clause (SR) for learning.

d Assessment

Two different but comparable versions of the picture description task adapted from Izumi and Bigelow (2000) were utilized to gauge the development of the participants’ knowledge of SRs. For each item, participants would see a picture and hear two simple declarative sentences which depicted this picture (e.g. The boy is Tom and He is cleaning the desk). The participants were then required to complete a sentence that identified the person/animate agent in the picture, with a prompt (e.g. The boy _______). To help them use English SRs, they were told to follow the instruction about the focus for each sentence: You should focus on what the boy is doing to the desk. The participants would not be allowed to combine the sentences they heard using and. This task consisted of 12 items. All words that were possibly used by participants in this task were from the word-list of their course book. The pilot test administered to another 10 learners suggested that there were no strange words to them in the learning and testing materials.

e Data scoring

All sentences produced by the participants were scored. Only the use of SRs that was targeted in the given item was scored by giving 1 point for a correct response and 0 for an incorrect one. A correct response means that the SR should contain the relative pronoun (i.e. that) and the object (i.e. the desk) as in the boy that is cleaning the desk is Tom. An incorrect response means that:

the relative pronoun was omitted (e.g. the boy is cleaning the desk is Tom);

the SR contained the subject (e.g. the boy he is leaning the desk is Tom);

two simple declarative sentences were conjoined by and (e.g. The boy is Tom and he is cleaning the desk);

two simple declarative sentences were separated with a stop (e.g. The boy is Tom. He is cleaning the desk).

Participants did not produce other types of grammatical SR forms such as The boy cleaning the desk is Tom. Errors regarding misspelling, articles, prepositions, tenses and so on were ignored.

2 Data analysis and discussion

As shown in Figure 2, all groups obtained zero scores in the pretest. After training, all experimental groups outperformed the controls in using SRs in the post and delayed posttest, while the controls’ performance remained stable. Specifically, both the ZF and BF groups achieved the mean scores of 5.7 and 5.3 in the posttest, respectively, while their mean scores decreased in the delayed posttest, with 4.0 and 3.9 for each.

Figure 2.

Participants’ mean scores of subject-extracted relative clauses (SRs) in the picture description task.

To test the effect of input frequency, the mixed-effects logistic regression analysis was carried out via the statistics package lme4 in the R environment (Barr et al., 2013; R Core Team, 2015), because the participants’ response (correct/incorrect) was as a binary variable. Frequency type (ZF/BF/CG) and test (pretest / posttest /delayed posttest) were fixed effects and centered with the Helmert coding to make interactions between the two fixed effects interpretable (UCLA Statistical Consulting Group, 2011). Participants, items, and by-item slopes for frequency were treated as random effect variables. The maximal random-effects structure for which the model would converge was employed (Barr et al., 2013). The two experimental groups at the pretest achieved zero scores, suggesting that all participants did not have the knowledge of English SRs. The control group at all tests achieved zero scores, revealing that the difference of the two experimental groups in all tests was induced by training instead of other extraneous variables. The inclusion of zero scores of the three groups at the pretest and zero scores of the control group at all tests lead to the failure of convergence for the logistic mixed-effects models, so they were excluded. A forward-fitting strategy was employed in model building. In the first model, only random effects were included. In subsequent models, single fixed effects and their interaction terms were added in turn. The effect of interaction between frequency and test was not significant. Model comparison using the ANOVA function suggested that the model that contained the two fixed effects but no interaction was best fitting:

Model<- glmer (Scores ~ frequently + test + (1|participant) + (1+ frequently|item)

As shown in Table 2, the effect of test (β = 0.299, SE = 0.069, Z = −4.314, p < .001) was significant, while the effect of frequency did not approach the significant level (β = 0.032, SE = 0.070, Z = −0.454, p = 0.650). This demonstrates that, (1) there were no differences regarding the participants’ production of SRs between the ZF and BF groups in all after-training tests, indicating that it did not make a difference whether the input was skewed toward a high token-frequency instance or balanced across all instances; and (2) participants in the ZF and BF groups achieved significantly higher scores in the posttest than in the delayed posttest, revealing the learning gains decreased substantially after a week.

Table 2.

Logistic mixed-effects modeling the experimental groups’ production of SRs after training.

Fixed effects	β	SE	Z	Pr(>Z)
(Intercept)	−0.331	0.112	−2.948	0.003
Frequency (ZF vs. BF)	−0.032	0.070	−0.454	0.650
Tests (posttest vs. delayed posttest)	−0.299	0.069	−4.314	< 0.001
Random effects	SD
Participants (intercept)	< 0.001
Items (intercept)	0.304
Frequency (slope)	0.013

Notes. SE = standard error. SD = standard deviation.

To recap, although the treatment had positive impact on the learning of SRs in general, both the ZF and BF groups behaved similarly in describing the pictures. This could be taken as additional evidence that Zipfian input did not differ from balanced input in boosting the learning of English SRs, as found by Dong and Zhang (2017).

VII Study 2

Study 2 was intended to answer Questions 2 and 3. We focused particularly on how input frequency interacted with the already learned SR to influence the development of ORs. When input frequency and language proficiency of the participants who had learned SRs were comparable to input frequency and language proficiency of those who had not learned SRs, it is possible to observe the interaction between input frequency and interference from the already learned SR via comparing the learning gains of the two groups.

1 Method

a Participants

The ZF, BF groups and CG from Study 1 were recruited. To obtain the baseline data, 40 third-year Chinese students that did not learn English SRs were included. They were from the same population as those in Study 1 and had no previous knowledge of the SR construction (See detailed information in Participant Section of Study 1). The 40 learners were evenly assigned to two groups, labeled as baseline ZF, and baseline BF, respectively. One-way ANOVA on the scores of their recent English midterm test suggested that they and the three groups from Study 1 were of the same proficiency (F(4, 95) = 1.249, p = .296). The biodata of the baseline groups are presented in Table 3.

Table 3.

Biodata of the two baseline groups.

Groups	n	Age	Male; Female	Midterm test scores
Groups	n	Age	Male; Female	Mean	SD
Baseline ZF	20	14–16	9;11	81.05	5.20
Baseline BF	20	14–16	13;7	82.90	5.78

Notes. ZF = Zipfian input; BF = balanced input.

b Training

Participants learned the OR with matched pictures. The OR and matched pictures were presented on slides. As with Study 1, the construction learning consisted of two phases (see Figure 3). On Phase 1, Picture A and the two sentences (e.g. The girl is her sister and The woman is hugging the girl) that described this picture occurred on the side. Participants looked at the picture and read the two sentences once after the second author. On Phase 2, Picture B and an OR that was made up of the two sentences describing Picture B (e.g. The girl that the woman is hugging is her sister) were presented. In Picture B, an arrow was provided to ensure that participants’ attention would be directed to the entity that would be modified by the OR. The ZF and baseline ZF groups received Zipfian input of ORs, the BF and baseline BF groups received balanced input of ORs, while the CG attended regular classes. The frequency and presentation order of all instances were the same as those in Study 1. The learning session lasted about 20 minutes.

Figure 3.

An example of a object-extracted relative clause (OR) for learning.

c Assessment

Two different but comparable versions of the picture description task were used to test the development of the participants’ knowledge of English ORs. For each item, learners would see a picture and hear two simple declarative sentences depicting this picture (e.g. The football is blue and The boy is playing the football). The participants were then required to complete a sentence that identified the patient in the picture, with a prompt like The football_______. The learners were also told to follow the instruction about the focus for each sentence (e.g. you should focus on the relationship between the football and the boy in the picture and not to directly combine the sentences you heard using and). This task contained 12 testing items.

d Data scoring

Only the OR produced by participants for the item was scored. A correct response was awarded 1 point, while an incorrect response 0 points. A correct response was defined as the OR that contained the relative pronoun (e.g. that, who) but did not contain the object (e.g. the football that the boy is playing is blue). If the relative pronoun (e.g. that) was omitted, it was also taken as correct. An incorrect response was identified as:

the OR contained the object (e.g. The football that the boy is playing it is blue);

two declarative sentences were conjoined by and (e.g. The football is blue and The boy is playing the football); and

two declarative sentences were separated with a stop (e.g. The football is blue. The boy is playing the football).

Other errors were ignored.

e Data collection

The procedures of data collection were the same as those in Study 1.

2 Data analysis and discussion

As shown in Figure 4, all groups achieved zero scores in the pretest. After training, the scores of four experimental groups became higher in the posttest (baseline ZF: 4.7; baseline BF: 4.80; ZF: 4.75: BF: 3.85) and the delayed posttest (baseline ZF: 3.75; baseline BF: 3.75; ZF: 3.70: BF: 2.75). It seems that the groups exposed to ZF did better than those exposed to BF in both the posttest and delayed posttest. To find whether there was significant difference in terms of OR use among the five groups across the three tests, the mixed-effects logistic regression analysis was performed, with the maximal random-effects structure for which the model would converge being used, where participants and testing items were treated as random effects, while frequency type, test, and the interaction between the two were treated as fixed effects. The data analyses consisted of three steps. First, we used the baseline data to test whether the input effects obtained in Study 1 could also be observed when learners learned English ORs without the previous experience of English SRs. Second, we used both baseline data and experimental data to examine whether the input effects detected based on the baseline data retained when learners learned English ORs with previous experience of English SRs. Third, we compared misuse of SRs by the ZF and BF groups in the picture description task to infer the role of different input in inhibiting the interference from previous experience of SRs.

Figure 4.

Participants’ mean scores of object-extracted relative clauses (ORs) in the picture description task.

a Testing input effect on OR learning via baseline data

Frequency type (Baseline ZF/Baseline BF/CG) and test (pretest / posttest / delayed posttests) were Helmert coded in order to make interactions interpretable. Procedures of data analysis were the same as those for Study 1. The two baseline groups in the pretest and the control group in all tests achieved zero scores, suggesting that all participants in this study did not have the knowledge of English ORs, and the difference of the two baseline groups in all tests were triggered by our treatment instead of other extraneous variables. The inclusion of zero scores of the two baseline groups at the pretest and zero scores of the control group at all tests lead to the failure of convergence of the mixed-effects logistic model, so we excluded them in the data analysis. Model comparison suggested that the model containing the two fixed effects but no interaction was best fitting, since the effect of interaction between frequency and test was not significant:

Model<- glmer (Scores ~ frequently + test + (1|participant) + (1+ frequently|item)

As shown in Table 4, the mixed-effects model yielded significant effect for test (β = −0.206, SE = 0.070, Z = −2.936, p = .003), but non-significant effect for input frequencies (β = 0.010, SE = 0.071, Z = 0.146, p = 0.884), demonstrating that the baseline ZF and BF groups did not differ in using the OR to describe pictures at all tests, and both groups did better at the posttest and delayed posttest. These finding are consistent with those obtained in Study 1, namely that ZF was not advantageous than BF in L2 acquisition of ORs. In the following part, we tested whether this trend also obtained in learners with previous experiences of English SRs.

Table 4.

Mixed-effects modeling the baseline groups’ production of ORs after training.

Fixed effects	β	SE	Z	Pr(>Z )
(Intercept)	−0.539	0.132	−4.092	< 0.001
Frequency (baseline ZF vs. baseline BF)	0.010	0.071	0.146	0.884
Tests (posttest vs. delayed posttest)	−0.206	0.070	−2.936	0.003
Random effects	SD
Participants (intercept)	0.027
Items (intercept)	0.383
Frequency (slope)	0.008

Notes. SE = standard error. SD = standard deviation.

b Testing input effect on OR learning via experimental data

Frequency type (baseline ZF/baseline BF/ZF/BF/CG) and test (pretest / posttest / delayed posttest) was Helmert coded. The procedures of data analysis were the same as those for Study 1. The baseline and experimental groups in the pretest and the control group in all tests achieved zero scores. Owing to the fact that the inclusion of zero scores of these groups caused the logistic mixed-effects model to fail to converge, we excluded them in the data analysis. Model comparison suggested that the model that only contained the two fixed effects was best fitting due to the fact that the effect of interaction between frequency and test was not significant:

Model<- glmer (Scores ~ frequently + test + (1|participant) + (1+|item)

As shown in Table 5, the effects for input frequency and test were significant. To be specific, all groups achieved higher scores in the posttest than in the delayed posttest (β = 0.310, SE = 0.091, Z = 3.428, p = .001). The participants who received ZF outperformed those who received BF after training (β = 0.310, SE = 0.091, Z = 3.428, p = .001), suggesting that Zipfian frequency is more beneficial than balanced frequency for learners with previous experience of English SRs to learn English ORs in both the short and long run. In addition, between-group comparisons suggested non-significant difference between the ZF and the baseline conditions (ZF vs. baseline ZF: β = −0.122, SE = 0.087, Z = −1.397, p = 0.162; ZF vs. baseline ZF: β = 0.098, SE = 0.087, Z = 1.124, p = 0.261). This seems to demonstrate that the ZF group were not influenced by their precious exposure to English SRs, while the BF group’s learning of ORs may be interfered by their exposure to English SRs. To further test this possibility, in the following section, we compared the misuse of SRs by the both experimental groups in the process where they did the picture description task.

Table 5.

Logistic mixed-effects modeling the experimental groups’ production of ORs after training.

Fixed effects	β	SE	Z	Pr(>Z)
(Intercept)	−0.661	0.142	−4.668	< 0.001
Frequency (ZF vs. baseline ZF)	−0.122	0.087	−1.397	0.162
Frequency (ZF vs. BF)	0.310	0.091	3.428	0.001
Frequency (ZF vs. baseline BF)	0.098	0.087	1.124	0.261
Tests (posttest vs. delayed posttest)	−0.214	0.051	−4.208	< 0.001
Random effects	SD
Participants (intercept)	0.041
Items (intercept)	0.455

Notes. SE = standard error. SD = standard deviation.

c Comparing misuse of SRs between the ZF and BF groups

We used the scoring scheme in Study 1 to identify the use of SRs by the ZF and BF groups in the picture description task.² The frequency of SR use is presented in Figure 5. Mann-Whitney U tests revealed that the ZF group did not differ from the BF in using SRs at the pretest (Z = 0.306, p = .760), but used less SRs at both the posttest (Z = −4.110, p < .001) and delayed posttests (Z = −2.089, p = .037), revealing that ZF could inhibit more interference from SR than did the BF in both the short and long term. Taken together, the analyses of learning gains of ORs and misuse of SRs revealed that learners exposed to BF were subject to the interference from previously learned SRs.

Figure 5.

Frequency of subject-extracted relative clause (SR) uses in the picture description task of Study 2.

The analysis of baseline data yielded the same pattern as that found in Study 1, demonstrating that ZF and BF did not differ in Chinese speakers’ learning of SRs and ORs. However, the analysis of experimental data and difference between baseline and experimental data suggest that the ZF group outperformed the BF group in all tests, the BF group achieved significantly lower scores than its baseline group, and the ZF group did the same as its baseline group. This may speak to the fact that the lower scores in the BF group was attributable to interference from the previously learned SR, while higher score in the ZF group could be interpreted as the inhibitory effect on the SR interference. That is, ZF was more effective than BF when learners learned ORs after they had learned SRs.

VIII General discussion

The present study examined the effect of the interaction between input frequency and learners’ prior knowledge on the acquisition of English-L2 relative clauses. Results are that (1) the effectiveness of ZF and BF at facilitating Chinese speakers’ acquisition of SRs and ORs did not differ when learners were not influenced by the related construction, and (2) ZF was more beneficial than BF for learners to learn English ORs when learners had previous exposure to English SRs. These results may point to the fact that ZF played a significant role in inhibiting constructional interference in L2 learning of English relative clauses.

The first two questions asked whether the effect of ZF and BF on the acquisition of English SRs and ORs differs. Results obtained from Study 1 and baseline data in Study 2 did not suggest the advantage of ZF over BF. Our findings were partly consistent with those from previous studies of English ditransitives (McDonough and Nekrasova-Becker, 2014; Year and Gordon, 2009), and Esperanto ditransitives (McDonough and Trofimovich, 2013). For this non-advantage of ZF, we ventured the following possible explanations. Compared with other constructions such as English VACs, SRs and ORs are much more abstract. It is difficult to identify a specific instance that can serve as a prototypical exemplar. The instance used in both ZF and BF may be of the same prototypicality in semantics. The role of the path-breaking instance assumed in ZF was therefore constrained. As reviewed previously, when the prototypicality of the skewed instance remained opaque as in McDonough and Nekrasova-Becker (2014), McDonough and Trofimovich (2013), and Year and Gordon (2009), the advantage of ZF would not be observed. Another possible account is that participants in our study were only exposed to 6 types of SR/OR instances. It was easy for the participants under both the ZF and the BF conditions to memorize individual instances with the help of visual images. Learners under all experimental conditions may be able to produce new items on the basis of their memories of the pattern of SRs/ORs, therefore diminishing the beneficial effect of ZF.

The way the target language is learned (implicit vs. explicit) may help learners develop different propensities in information processing. As argued by McDonough and Trofimovich (2013), learners accustomed to explicit instruction prefer to analyse the sentence structure while this is not the case for those exposed to implicit instruction. The participants in our two studies were Chinese speakers who had learned English explicitly in classroom settings. Chances were that our participants equipped with more analytic mindsets were used to analysing grammatical features. If so, learners exposed to ZF and BF could find features of the SR/OR with equal possibility. This may result in the null difference between ZF and BF. Additionally, the pattern used in both SR and OR was V+O (i.e. that is chasing the man) and S+V (i.e. that the woman is hugging), two varieties of S+V+O that the participants were familiar with, so it is easy for them to detect the pattern.

One point needs acknowledging is that even if SRs and ORs varied in complexity (or difficulty), the participants in Study 1 and the baseline group in Study 2 displayed similar patterns. Extant evidence seems to suggest that the OR is more difficult than the SR due to the fact that the extracted nominal phrase (the Subject) in the SR is near to the verb while the extracted nominal phrase (the Object) in the OR is relatively far from the verb (Mecklinger et al., 1995), as shown in The girl that is hugging the woman is her sister and The girl that the woman is hugging is her sister. The participants would spare less cognitive resources in processing the SR than the OR because the girl in the former is closer to hugging than in the latter. However, this supposed effect was not observed in our study. The possible account is that both SRs and ORs used in this study took the form of SV or VS, our participants, as their instructor commented, had learned the five English basic structures (see Note 3), so they were developmental ready for these target constructions. We postulated that it was this ‘readiness’ that made our participants experience the same pattern in learning the SR and OR. It is important to note, however, that previous L2 research detected that learners exposed to BF outperformed those exposed to ZF (McDonough and Nekrasova-Becker, 2014; McDonough and Trofimovich, 2013; Year and Gordon, 2009). In Year and Gordon (2009), BF contained 40 instances of English ditransitives, in McDonough and Nekrasova-Becker (2014) it included 20 items, and in McDonough and Trofimovich (2013) it comprised 24 items. However, the participants in our study learned only 12 instances of SRs/ORs. It is likely that the facilitative role of BF would be enhanced when the type frequency increases. The small number of types of instances in our two studies may render the role of BF less prominent.

Results of the experimental data in Study 2 suggest that the previously learned SRs exerted negative influence on L2 learning of ORs. Specifically, learners exposed to BF did worse than the corresponding baseline group, while learners exposed to ZF achieved the same scores as those in the related baseline group in the test. Analysis of misuse of SRs also indicated that learners exposed to BF generated more errors of SRs than those exposed to ZF in Study 2. This means that BF was less effective than ZF to help learners hinder the influence of the previously learned SRs. Remember that the participants learned ORs after they finished the delayed posttest for SRs. Their scores of SRs at the delayed posttest were significantly lower than the scores at the posttest for each group, suggesting that the participants’ ability to use SRs to depict the picture under both ZF and BF conditions had decreased as time went by. When encountering the OR, they were supposed to distinguish the OR from the already learned SRs. As Clark and Clark (1979) predict, if two constructions occur in similar contexts, learners may consider them to be identical in meaning. If our participants had learned a form that expressed a meaning of modification, then it is difficult for them to use another form to express this meaning. Following this, the SR would block the learning of the OR in the initial learning stage. This SR interference was more prominent in the BF group than in the ZF group. Although ZF and BF were found to play similar roles in SR learning in Study 1, ZF provided high token-frequencies of ORs, which made it easier for learners to remember the high-frequency OR instance, thus releasing more cognitive resources than those exposed to BF in Study 2. Chances were that the ZF group was more likely to detect the difference between SR and OR with the help of the skewed instance. From the perspective of language processing, ZF potentially helped learners identify the cues for the OR with low-variance input (Crossley et al., 2013; Elio and Anderson, 1984). In doing the picture description task, they tended to use the ‘anchor’ as a benchmark, making the SR interference effectively reduced (Wilson, et al., 1996; Zhang and Dong, 2016). A scrutiny of the OR used by the two groups suggests that the BF group used more erroneous ORs like the football that the boy played it is blue than did the ZF group at both posttest and delayed posttest. This supports the account that it was more difficult for the BF group in Study 2 to distinguish SRs from ORs. The BF group tended to treat the football that the boy is playing it and the boy that is playing the football as the same. Additionally, the correct use of OR for the ZF group and its baseline group was the same, suggesting that the ZF group was less influenced by SR in the learning of ORs. In this sense, BF was less effective than ZF to hinder the interference of SRs.

The present study has pedagogical implications as well. When the target construction is too abstract like SRs and ORs that may not have specific prototypical instances, learning gains triggered by ZF and BF may not differ. Either of skewed or balanced input could be used when L2 learners are learning abstract constructions like relative clauses. Teachers should be aware of whether the construction targeted is similar to the learners’ previously-learned construction. If so, in order to avoid potential negative influence from the already learned construction, teachers are encouraged to initially provide low-variance input wherein learners can anchor a given instance when learning other instances, so as to maximize the possibility of form-meaning mappings for the target construction. A practical concern for L2 teaching is how to help learners avoid negative influence from similar structures (Larsen-Freeman, 2013). Language instructors are therefore expected to find effective ways like manipulating type and token frequencies as well as the presentation order of instances in input to help learners escape from such interference. In short, in order to maximize the learning potential in classroom settings, teachers are encouraged to pay attention to the interaction among factors like frequency, presentation order, and influence from other similar constructions, which are of crucial importance for L2 development.

IX Conclusions

This study yielded findings that ZF and BF generated similar gains in the Chinese speakers’ acquisition of English SRs/ORs when the two constructions were learned independently, and that ZF was more effective than BF to hinder the interference from SRs in the process where ORs were learned. That is, when considering the previously-learned constructions, ZF is more effective for learners to learn English ORs. However, the present study only tested the effect of different input frequencies on L2 acquisition of English SRs and ORs, which constrains its generalizability. It is therefore necessary to exercise caution in applying our findings to other constructions, languages, and language learners. One shortcoming in the experimental design of the present two studies that needs acknowledging is that one group of students receiving some kind of dummy training should be recruited in order to control for task effects. It is possible that L2 growth we observed may be contributed to the task effects instead of training with matched pictures.³ This should deserve attention in future studies. Another shortcoming pertains to the unnaturalness of a few test items designed to induce the participants’ response in the second study (e.g. The boy is playing the football. The dog is biting his uncle. The girl is combing the hair). Although these sentences were selected from the participants’ test book and the participants did not report that these sentences were unnatural to them in the pilot test, the use of these test items might influence the observed pattern in our study.

Supplemental Material

appendix- – Supplemental material for Input frequency and construction interference interactions in L2 development

Supplemental material, appendix- for Input frequency and construction interference interactions in L2 development by Xiaopeng Zhang and Xiaoli Dong in Second Language Research

Footnotes

Acknowledgements

We would like to thank the students who took part in this research. We are deeply grateful to Roumyana Slabakova, and four anonymous Second Language Research reviewers for their extensive comments and suggestions.

Declaration of Conflicting Interest

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was supported by a grant from the National Social Science Foundation of China (Grant Number: 15CYY018) awarded to Xiaopeng Zhang.

Supplemental material

Supplementary material for this article is available online.

Notes

ORCID iD

Xiaopeng Zhang

References

Abbot-Smith

Behrens

(2006) How known constructions influence the acquisition of other constructions: The German passive and future constructions. Cognitive Science 30: 995–1026.

Ambridge

Bidgood

Twomey

, et al. (2015) Preemption vs. entrenchment: Towards a construction general solution to the problem of the retreat from verb argument structure overgeneralization. PLoS ONE 10: 1–20.

Barr

Levy

Scheepers

Tily

(2013) Random effects structure for confirmatory hypothesis testing: Keep it maximal. Journal of Memory and Language 68: 255–78.

Bybee

(2008) Usage-based grammar and second language acquisition. In: Robinson

Ellis

(eds) Handbook of cognitive linguistics and second language acquisition (pp. 216–36). New York: Routledge.

Cai

(2006) Accessibility hierarchy hypothesis and acquisition of English relative clauses. Modern Foreign Language 29: 56–65.

Casenhiser

Goldberg

(2005) Fast mapping between a phrasal form and meaning. Developmental Science 8: 500–08.

Clark

(1979) When nouns surface as verbs. Language 55: 767–811.

Crossley

Subtirelu

Salsbury

(2013) Frequency effects or context effects in second language word learning. Studies in Second Language Acquisition 35: 727–55.

Dong

Zhang

(2017) Effects of different frequency distributions on L2 development of two English constructions. Modern Foreign Languages 40: 540–51.

10.

Eckman

(1984) Universals, typologies and interlanguage. In: Rutherford

(ed) Language universals and second language acquisition (pp. 79–105). Amsterdam / Philadelphia, PA: John Benjamins.

11.

Elio

Anderson

(1984) The effects of information order and learning mode on schema abstraction. Memory and Cognition 12: 20–30.

12.

Ellis

(2002) Frequency effects in language processing: A review with implications for theories of implicit and explicit language acquisition. Studies in Second Language Acquisition 24: 143–88.

13.

Ellis

(2006a) Language acquisition as rational contingency learning. Applied Linguistics 27: 1–24.

14.

Ellis

(2006b) Cognitive perspectives on SLA: The associative-cognitive CREED. AILA Review 19: 100–21.

15.

Ellis

Cadierno

(2009) Constructing a second language. Annual Review of Cognitive Linguistics 7: 111–290.

16.

Ellis

Collins

(2009) Input and second language acquisition: The roles of frequency, form, and function. Introduction to the special issue. Modern Language Journal 93: 329–35.

17.

Ellis

Ferreira-Junior

(2009) Construction learning as a function of frequency, frequency distribution, and function. Modern Language Journal 93: 370–85.

18.

Ellis

Sagarra

(2011) Learned attention in adult language acquisition: A replication and generalization study and meta-analysis. Studies in Second Language Acquisition 33: 589–624.

19.

Ellis

O’Donnell

Römer

(2013) Usage-based language: Investigating the latent structures that underpin acquisition. Language Learning 63: 25–51.

20.

Fulga

McDonough

(2016) The impact of first language background and visual information on the effectiveness of low-variability input. Applied Psycholinguistics 37: 265–83.

21.

Gass

(1979) Language transfer and universal grammatical relations. Language Learning 29: 327–44.

22.

Goldberg

(2006) Constructions at work: The nature of generalization in language. Oxford: Oxford University Press.

23.

Goldberg

Casenhiser

(2008) Construction learning and second language acquisition. In: Robinson

Ellis

(eds) Handbook of cognitive linguistics and second language acquisition (pp. 197–215). New York: Routledge.

24.

Goldberg

Casenhiser

Sethuraman

(2004) Learning argument structure generalizations. Cognitive Linguistics 15: 289–316.

25.

Goldberg

Casenhiser

White

(2007) Constructions as categories of language. New Ideas in Psychology 25: 70–86.

26.

Izumi

Bigelow

(2000) Does output promote noticing and second language acquisition? TESOL Quarterly 34: 239–78.

27.

Keenan

Comrie

(1977) Noun phrase accessibility and Universal Grammar. Linguistic Inquiry 8: 63–99.

28.

Kidd

Lieven

Tomasello

(2006) Examining the role of lexical frequency in the acquisition and processing of sentential complements. Cognitive Development 21: 93–107.

29.

Krajewski

Siebenborn

A-K

Lieven

EVM

(2011) The shape of frequency distribution and novel construction learning. Unpublished paper presented at the Boston University Conference on Language Development, Boston, MA, USA.

30.

Larsen-Freeman

(2013) Transfer of learning transformed. Language Learning 63: 107–29.

31.

McDonough

Nekrasova-Becker

(2014) Comparing the effect of skewed and balanced input on English as a foreign language learners’ comprehension of the double-object dative construction. Applied Psycholinguistics 35: 419–42.

32.

McDonough

Trofimovich

(2013) Learning a novel pattern through balanced and skewed input. Bilingualism: Language and Cognition 16: 654–62.

33.

Mecklinger

Schriefers

Steinhauer

Friederici

(1995) Processing relative clauses varying on syntactic and semantic dimensions: An analysis with event-related potentials. Memory and Cognition 23: 477–94.

34.

Morris

Cottrell

Elman

(2000) A connectionist simulation of the empirical acquisition of grammatical relations. In: Wermter

Sun

(eds) Hybrid neural symbolic integration (pp. 175–93). Berlin: Springer.

35.

Nakamura

(2012) Input skewedness, consistency, and order of frequent verbs in frequency-driven second language construction learning: A replication and extension of Casenhiser and Goldberg (2005) to adult second language acquisition. International Review of Applied Linguistics in Language Teaching 50: 31–67.

36.

Pajak

Fine

Kleinschmidt

Jaeger

(2016) Learning additional languages as hierarchical probabilistic inference: Insights from first language processing. Language Learning 66: 900–44.

37.

R Core Team (2015) R: A language and environment for statistical computing. Vienna: R Foundation for Statistical Computing.

38.

Richards

(1971) A non-contrastive approach to error analysis. English Language Teaching Journal 25: 204–19.

39.

Schachter

Celce-Murcia

(1977) Some reservations concerning error analysis. TESOL Quarterly 11: 441–51.

40.

Scovel

(2001) Learning new languages: A guide to second language acquisition. Boston, MA: Heinle and Heinle.

41.

Tomasello

(2003) Constructing a language: A usage-based theory of language acquisition. Cambridge, MA: Harvard University Press.

42.

UCLA Statistical Consulting Group (2011) R library: Contrast coding systems for categorical variables. Los Angeles, CA: UCLA Statistical Consulting Group. Available at: https://stats.idre.ucla.edu/r/library/r-library-contrast-coding-systems-for-categorical-variables (accessed July 2018).

43.

Wilson

Houston

Etling

Brekke

(1996) A new look at anchoring effects: Basic anchoring and its antecedents. Journal of Experiment Psychology: General 125: 387–402.

44.

Wulff

Ellis

Römer

Bardovi-Harlig

LeBlanc

(2009) The acquisition of tense–aspect: Converging evidence from corpora and telicity ratings. Modern Language Journal 93: 354–69.

45.

Year

Gordon

(2009) Korean speakers’ acquisition of the English ditransitive construction: The role of verb prototype, input distribution, and frequency. Modern Language Journal 93: 389–417.

46.

Zhang

Dong

(2016) Revisiting Zipfian frequency: L2 acquisition of English prenominal past participles. Modern Language Journal 100: 404–27.

47.

Zhang

(2014) The effect of frequency distribution of language input on the acquisition of English phrasal construction by Chinese EFL learners. Modern Foreign Language 37: 78–88.

48.

Zipf

(1935) The psycho-biology of language: An introduction to dynamic philology. Cambridge, MA: The MIT Press.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.84 MB