A Data-Driven Critical Review of Second Language Acquisition in the Past 30 Years

: This study aims to provide a comprehensive and data-driven review of the knowledge domain of second language acquisition (SLA) and pedagogy in the past 30 years. Using knowledge domain visualization techniques, the study ﬁrst provides a review of SLA at the disciplinary level. It then identiﬁes the major research areas and current research frontiers in the SLA research landscape based on high-quality data retrieved from Web of Science (WoS) databases. The study provides useful references for future research and pedagogy in the ﬁeld in which literature reviews employing scientometric methodology and driven by data, such as the present one, are rare, and thus, are much in need of supplement views produced by traditional literature reviews.


Introduction
Second language acquisition (SLA), or L2 (language 2) acquisition, studies the learning process of any language other than a person's first language. Although strictly speaking, the word acquisition does not refer to the conscious learning or the teaching of a language [1], and the phrase second language differs from foreign language [2], researchers have used both word pairs-learning/acquisition and second language/foreign language-as synonyms, and they have also extended SLA to include the learning process of the third, fourth, or subsequent languages [3][4][5]. However, SLA has been preferred by most researchers in this discipline.
VanPatten and Benati suggest that two works, Corder's The Significance of Learners' Errors [6] and Selinker's Interlanguage [7], mark the beginning of the modern study of SLA as a scientific discipline [8] (pp. [2][3][4][5]. Corder's systematic analysis of learner errors provides empirical evidence on how a language is learned [6]. Selinker defines interlanguage as a learner language system with its own right and a complete set of rules which is different from the languages that learners already know [7]. Both works are regarded as facilitating the development of SLA from a branch of applied linguistics into an academic discipline built upon linguistics, psychology, cognitive psychology, sociology, sociolinguistics, and education [4,9]. Approaches and theories of SLA have been introduced and extensively elaborated upon in a large amount of works, including but not limited to Ellis [2], Gass and Selinker [10], Larsen-Freeman and Long [11], and Ortega [5], and VanPatten and Williams [12]. Some of the prominent areas included in the works are Contrastive Linguistics Hypothesis [13], Error Analysis [6], Interlanguage Theory [7], Monitor Theory [1,14], Zone of Proximal Development (ZPD) [15], Universal Grammar and Minimalist Program [16,17], Corpus Linguistics [18][19][20][21], and Cognitive Approach [2]. The following is a brief introduction to some of the most important works in the SLA field. The introduction will be arranged according to a chronological order.

1.
Acquisition/Learning Hypothesis. It proposes that "adults have two distinct and independent ways of developing competence in a second language" [1] (p. 65). Acquisition refers to the subconscious way, while learning is the conscious way.

2.
Natural Order Hypothesis. It claims that if learners learn a second language through natural means, their learning order will show a high degree of predictability. 3.
Monitor Hypothesis. It states that monitoring or editing is the only purpose for our conscious to learn the rules of a language. 4.
Input Hypothesis. In contrast to error analysis, this hypothesis claims that comprehensible input is the most important factor for learners to acquire a second language. 5.
Affective Filter Hypothesis. It states that in addition to comprehensible input, the acquisition of a language depends on the learner's emotions, needs, attitude, and motivations for second language learning.
Despite its many deficiencies, Monitor Theory has played a significant role in promoting the development of SLA theory and teaching and has a higher or lower impact on the development of a number of other important hypotheses such as Swain's Comprehensive Output Hypothesis [27] and Schmidt's Noticing Hypothesis [28]. In particular, the Input Hypothesis and Affective Filter Hypothesis have both had an important practical significance to language teaching and practice [11].
Corpus Linguistics, an interdisciplinary field, has attracted a great deal of attention since the 1980s [18][19][20][21]. It studies the collection, storage, processing and statistical analysis of natural language texts in order to engage in linguistic research with objective and informative linguistic evidence provided by large-scale corpora [29]. According to Chen [30], there are 5 major research streams of corpus linguistic studies in SLA (i.e., lexical acquisition, language acquisition, cognitive studies, vocabulary teaching and learning, and corpora) in the classroom. Although corpus linguistics has its own unique theoretical system and methodology, it is not yet regarded as a distinct branch of linguistics, but rather a research method [31,32].
Modern SLA research has turned to cognitive approaches, and the most important model in respective a cognitive approach is Ellis's [2] computational model [8]. According to Ellis, the learning process starts with a first stage where certain features of the language are deposited in the learner's short-term memory through input [2]. In the next stage, some of the deposits are converted into second-language knowledge and stored in the learner's long-term memory. In the third stage, the learner uses the second language knowledge to produce output. Cognitive theories attempt to explore the mental representations and underlying process of the above 3 stages.
In addition, there are linguistic and sociocultural approaches to SLA (i.e., Chomsky's [16] Universal Grammar (UG) and Vygotsky's [15] Zone of Proximal Development (ZPD) [8]). The central idea of UG is that the human brain has a congenitally specific structure or attribute, that is, a language acquisition mechanism, which is the internal reason why humans learn to use language [16]. Within UG, a major mode of inquiry has developed into the Minimalist Program, basically, a conceptual framework that guides the inquiry of linguistic theory [17]. On the other hand, ZPD is regarded as a unique sociocultural approach to SLA studies. It suggests that learners who interact with more advanced target language users will learn the language more efficiently than those who learn the language independently [15].
It is clear that SLA is closely related to many social-science fields, such as linguistics, psychology, psycholinguistics, pragmatics, education, sociology, sociolinguistics, and cognitive psychology. In the future, it will have a tendency to integrate with the natural sciences such as cognitive science, neuroscience, and neurobiology. It is expected that new research methods, such as mathematics, statistical analysis, and computational analysis, will be widely used in SLA research [8].
In the multidisciplinary field of SLA study, regular reviews of previous research are considered instrumental because the advancement of knowledge relies on an accumulation of individual theoretical and empirical studies [30]. However, previous reviews of SLA research [3,8,[33][34][35][36][37][38] have been mainly qualitative in nature. They often emphasized a few research themes depending on the author's preferences, and they might not capture emerging SLA developments in a timely manner. For example, Archibald et al. [3] looked at content-based learning; Hucking and Coady [34], the vocabulary acquisition; and, Weinert [38], the role of formulaic language. Although a few studies were data-driven, they were also focused on smaller SLA areas, (e.g., the use of corpora in SLA by Chen [30]). It has been argued that SLA lacks a clearly-defined research scope, and comprehensive literature reviews are much in need to facilitate the dialogue between researchers who take different research perspectives and methods [8].
In recent years, a number of bibliometric studies have emerged in the neighboring fields of SLA that provide fresh, comprehensive views of their respective fields of study [30,[39][40][41]. For example, Lei and Liu identified 165 of the most popular research topics discussed by 10,028 articles published in 42 Social Science Citation Index journals related with Applied Linguistics [41]. They found that most of the frequently discussed topics in applied linguistics remained popular between 2005 and 2016, including communicative competence, academic discourse, vocabulary acquisition, discourse analysis, corpus-based studies, etc. While some topics related to socio-cultural issues (e.g., language policy, multilingualism, and ideology) experienced a significant increase in interest, the focus on topics concerning phonology, grammar, generative linguistics decreased significantly. In addition, neither a significant increase nor a significant decrease was found in traditional linguistic research topics such as collaborative writing, education policy, interactional feedback, and universal grammar. They further argued that the research trend of applied linguistics tended to turn to "sociocultural and language policy issues" from "some formal linguistic issues, such as phonology and syntax" [41] (p. 9). These findings from applied linguistics are of great importance and interest to the current study, and they are worth further exploration using SLA data. There were also bibliometric studies that thoroughly examined the publication output, trends and collaboration of authors in the academic disciplines of linguistics [42], world Englishes [43], and second language writing [44]. These studies show that the bibliometric approach is powerful in revealing research trends in language-related fields with findings that qualitative analysis cannot yield.
In addition, in her bibliometric study of corpus-based studies of second/foreign language, Chen [30] identified 4 research trends in corpus-based SLA studies including the writing of second language learners, the construction of corpora, vocabulary teaching and learning, and the production process of both first language and second language learners. However, this study was quite specific, and it provided little in the way of information on the development of SLA as a field of study.
Given the advantage of a bibliometric approach to provide a more comprehensive examination of the literature in a discipline and the lack of a discipline-specific bibliometric review in SLA, this study used CiteSpace [45], a bibliometric visualization software package, to analyze the bibliometric information in the academic papers published from 1987 to 2018 in journals indexed by Web of Science (WoS) databases. The wide range of data sources and the bibliometric visualization method provided a complete and intuitive overview of the research areas, development and evolution process, and research hot-spots in SLA. The specific objectives of the study were to analyze the location of SLA in the broad scientific knowledge domain; establish the research areas in SLA through co-citations analysis; and identify landmark articles and determine the emerging trends in SLA research. The study is of significance for the researcher if only to better understand and grasp the primary research directions in this field.

Materials and Methods
The methodological process of this study included four stages: (1) retrieval of bibliographic data; (2) dual-mapping overlay to map SLA at the disciplinary level; (3) clusters analysis based on bibliographic information of co-cited articles to identify research areas and research frontiers (4) identification of pivotal works through burst detection, and (5) a timeline analysis and burst detection on co-cited keywords to identify important keywords.

Data Collection
A total of 11,381 references were obtained. The reference data used in this paper were retrieved from the Web of Science databases using the following retrieval strategies:

•
The subject was set to the following: (second language acquisition) OR (L2 acquisition) OR (second language learn*) OR (L2 learn*) OR (foreign language acquisition) OR (foreign language learn*); • The publication period was set from "between year 1900 and 2018" (Retrieved date 10 January 2018); • Reference type was set to "article or proceedings paper"; language, "English"; • Data category was refined by the subjects of "linguistics" or "education educational research" or "language linguistics".

Data Analysis
CiteSpace provides a range of analytical tools to detect thematic patterns and emerging trends [45]. The following analytical tools were used in this study.

Dual-Mapping Overlays
The term 'dual-map' refers to the citing and cited component maps in the overall visualization. Given a set of articles, an overlay of the set visualizes the disciplinary concentrations of these articles and how they connect to various regions on the global map through their citation links [46].

Co-Citation Analysis
Citespace co-citation or co-occurrence analyses can be applied to different factors, such as cited reference, co-word, cited author, cited journal and so on. By studying these clusters and the relationships between these clusters in visual maps, some valuable information can be drawn.
In this study, we employed a co-citation analysis of references. The visual maps generated by Citespace were composed of nodes and links, respectively representing individual articles and co-citation between these articles. According to C. Chen's definition, the higher citation and centrality the node has, the larger impact the node has on the co-citation map [47].

Co-citation Timeline View
One of the main purposes of detecting emerging trends is to understand the flow of information among research areas: which areas contribute knowledge, and which areas are "borrowers" of knowledge. For this purpose, it is useful to track scientific paradigms as a function of time. Timeline views in CiteSpace II simultaneously show cited articles and citing times in order to highlight the mapping between a research front and its intellectual base [45].

The Location of SLA in Scientific Knowledge Domain
In order to display the maps of SLA at the disciplinary level, we employed the dual-map overlay technology [46] provided by CiteSpace to produce an easy-to-interpret representation of citations made by a wide variety of publications included in our dataset. Figure 1 shows the patterns of SLA articles published between 1987 and 2018 against the global map of scientific literature that includes over 10,000 scientific journals reflecting a wide range of disciplines.

Co-Citation Analysis
Citespace co-citation or co-occurrence analyses can be applied to different factors, such as cited reference, co-word, cited author, cited journal and so on. By studying these clusters and the relationships between these clusters in visual maps, some valuable information can be drawn.
In this study, we employed a co-citation analysis of references. The visual maps generated by Citespace were composed of nodes and links, respectively representing individual articles and cocitation between these articles. According to C. Chen's definition, the higher citation and centrality the node has, the larger impact the node has on the co-citation map [47].

Co-citation Timeline View
One of the main purposes of detecting emerging trends is to understand the flow of information among research areas: which areas contribute knowledge, and which areas are ''borrowers'' of knowledge. For this purpose, it is useful to track scientific paradigms as a function of time. Timeline views in CiteSpace II simultaneously show cited articles and citing times in order to highlight the mapping between a research front and its intellectual base [45].

The Location of SLA in Scientific Knowledge Domain
In order to display the maps of SLA at the disciplinary level, we employed the dual-map overlay technology [46] provided by CiteSpace to produce an easy-to-interpret representation of citations made by a wide variety of publications included in our dataset. Figure 1 shows the patterns of SLA articles published between 1987 and 2018 against the global map of scientific literature that includes over 10,000 scientific journals reflecting a wide range of disciplines. In Figure 1, the left side is the distribution of journals (citing journals) in which a source article is published of the citing articles, which represents the main subject to which the SLA belongs; and, the right side corresponds to the distribution of journals (cited journals) in which a reference is published, which represents the main reference SLA subjects. Colored curves represent paths of references, originating from the map on the left and pointing to the cited-journals map on the right. The positions of the starting and ending points of these curves tell us about how an article has built upon previous work because both citing and cited maps are divided into a number of thematic areas and each position on the map belongs to one of the given areas. In Figure 1, the left side is the distribution of journals (citing journals) in which a source article is published of the citing articles, which represents the main subject to which the SLA belongs; and, the right side corresponds to the distribution of journals (cited journals) in which a reference is published, which represents the main reference SLA subjects. Colored curves represent paths of references, originating from the map on the left and pointing to the cited-journals map on the right. The positions of the starting and ending points of these curves tell us about how an article has built upon previous work because both citing and cited maps are divided into a number of thematic areas and each position on the map belongs to one of the given areas. Figure 1 shows that the research literature of SLA primarily appears in the broad area near the left bottom on the citing map with the label of Psychology, Education, and Health. It is quite reasonable that articles in our dataset are mainly published in journals belonging to this category because it is the major discipline in social sciences and journals in this discipline usually publish studies of SLA. Citation curves that originate from the primary area on the left are split into a few major streams and point to numerous areas in the cited map on the right, indicating that articles in SLA journals cite distinct groups of journals, which include Psychology, Education, Health/Mathematical, Mathematics, Mechanics/Health, Nursing, Medicine/Systems, Computing, Computer/History, Philosophy, Records/Veterinary, Animal, Parasitology/Molecular, Biology, Genetics, etc.
In the early 1990s, when compared to other social sciences, SLA research was then a relatively new field, mostly borrowing from LI studies, pedagogical studies and methods of other related disciplines [48,49]. Gass found that most of the interaction between SLA and the other disciplines was unidirectional from other disciplines to SLA, which was not healthy for SLA development [48]. Nowadays, the results of our study show that the major publishers of SLA studies are journals concentrated in the category of "Psychology, Education, and Health"(the left side of the map in Figure 1), suggesting that SLA research is not much needed by disciplines of other knowledge categories. On the other hand, SLA research has greater "needs" from other disciplines as shown on the right side of the map in Figure 1. This indicates that the influence of SLA on other disciplines may not have increased much over the years. One obvious reason is that SLA is a more specific and narrower research field under the broader research disciplines of linguistics and education; and, SLA researchers will naturally "need" theories and studies of other established research disciplines such as psychology, mathematics, and computing. As pointed out by Gass, whether a discipline is "needed" by other disciplines is a measurement of the discipline's integration with the academic community [48]. In this regards, a stronger interaction between SLA and other disciplines will lead to the contribution to the integration of the broader linguistics and education disciplines with other disciplines.
As a complement to this disciplinary view of SLA research, we also produced a geographic citation map as shown in Figure 2.
Publications 2018, 6, x FOR PEER REVIEW 6 of 29 Figure 1 shows that the research literature of SLA primarily appears in the broad area near the left bottom on the citing map with the label of Psychology, Education, and Health. It is quite reasonable that articles in our dataset are mainly published in journals belonging to this category because it is the major discipline in social sciences and journals in this discipline usually publish studies of SLA. Citation curves that originate from the primary area on the left are split into a few major streams and point to numerous areas in the cited map on the right, indicating that articles in SLA journals cite distinct groups of journals, which include Psychology, Education, Health/Mathematical, Mathematics, Mechanics/Health, Nursing, Medicine/Systems, Computing, Computer/History, Philosophy, Records/Veterinary, Animal, Parasitology/Molecular, Biology, Genetics, etc.
In the early 1990s, when compared to other social sciences, SLA research was then a relatively new field, mostly borrowing from LI studies, pedagogical studies and methods of other related disciplines [48,49]. Gass found that most of the interaction between SLA and the other disciplines was unidirectional from other disciplines to SLA, which was not healthy for SLA development [48]. Nowadays, the results of our study show that the major publishers of SLA studies are journals concentrated in the category of "Psychology, Education, and Health"(the left side of the map in Figure 1), suggesting that SLA research is not much needed by disciplines of other knowledge categories. On the other hand, SLA research has greater "needs" from other disciplines as shown on the right side of the map in Figure 1. This indicates that the influence of SLA on other disciplines may not have increased much over the years. One obvious reason is that SLA is a more specific and narrower research field under the broader research disciplines of linguistics and education; and, SLA researchers will naturally "need" theories and studies of other established research disciplines such as psychology, mathematics, and computing. As pointed out by Gass, whether a discipline is "needed" by other disciplines is a measurement of the discipline's integration with the academic community [48]. In this regards, a stronger interaction between SLA and other disciplines will lead to the contribution to the integration of the broader linguistics and education disciplines with other disciplines.
As a complement to this disciplinary view of SLA research, we also produced a geographic citation map as shown in Figure 2.  Figure 2 gives the general distribution patterns of and network relations among the references in our dataset. The largest sources of publication in SLA are located in 3 major geographical regions: East Asia/Australasia, North America, and Europe. Each of the color links, which may consist of hundreds of thinner links, represents a citation relationship across the continents. At the global level, the citation relationship between East Asia/Australasia and North America is equally as strong as  Figure 2 gives the general distribution patterns of and network relations among the references in our dataset. The largest sources of publication in SLA are located in 3 major geographical regions: East Asia/Australasia, North America, and Europe. Each of the color links, which may consist of hundreds of thinner links, represents a citation relationship across the continents. At the global level, the citation relationship between East Asia/Australasia and North America is equally as strong as that between Europe and North America. However, the citation relationship between East Asia/Australasia and Europe is relatively weak, shown by the fewer color links. Since statistics of research and publication output are available from almost all major academic databases; statistical information of the links has been left out in this study.

The Major Areas of Research in SLA
To find out the major areas of research consistent with SLA, we carried out a co-citation analysis on the collection of publications in our dataset. After repeated testing, the synthesized network in Figure 3 was generated, containing 250,296 references. The map generated by this study produced a clearer structure with a better explanatory power (Modularity 1 = 0.8119; Silhouette 2 = 0.6759) when the following parameters were set: The rest settings were left as default.
The clustering function was performed by choosing 'T' as the labeling source and the log-likelihood ratio (LLR) as the method. The results returned 20 knowledge clusters, 11 of which were major clusters labeled from #0 to #10. The cluster labels were generated automatically based on the LLR clustering algorithm. In order to increase the aesthetical effect of the mapping, we increased the transparency of nodes and links so that they were only shown in the background.
According to Chen [45], Modularity Q values between 0.4-0.8 are acceptable. The network in Figure 3 has a Modularity Q value of 0.8119 and a Mean Silhouette value of 0.6759, which are considered very high, suggesting that the research areas in SLA are clearly defined in terms of co-citation clusters.
The different colors of the areas indicate the time when co-citation links in those areas appeared for the first time. Areas in blue were generated earlier than areas in green. Areas in yellow were generated after the green areas and so on. Consequently, the two blue clusters (Clusters #1 and #3) are the oldest ones, and yellow clusters (#4, #5, and #9) are the most recent clusters. The most prominent cluster (Cluster #0) together with Clusters # 8 and #10 are relatively older ones, and Clusters #2, #6, and #7 are relatively recent ones. As CiteSpace's clustering algorithm puts each topic into a single cluster instead of multiple clusters, the appearance of a topic in one period does not necessarily mean that this topic has not been popular in the past. For example, Cluster #9, which includes topics of reading comprehension and vocabulary, is suggested as one of the most recent clusters. In fact, reading comprehension and vocabulary have been found to be long-lasting popular research topics both in SLA and in applied linguistics [30,41]. Similarly, Cluster #7, which includes topics of usage-based approaches and working memory, is regarded as one of the relatively recent clusters that spanned from 2003 to 2010. However, usage-based approaches are still gaining an increasing level of current interest in recent years [41].
The above explanation only gives a brief overview of the major research areas in SLA from 1987 to 2018. The detail information of each major cluster is further provided in Appendix B, in which 1 Modularity value measures the extent to which a network can be decomposed to multiple components, or modules 2 The silhouette value of a cluster measures the quality of a clustering configuration. Its value ranges between −1 and 1. The highest value represents a perfect solution. the cluster size, publishing year, and cluster labels generated by other clustering algorithms (LSI and MI) are listed. The cluster labels generated by LSI and MI algorithms provided alternative views on the content of the clusters. However, they were too vague and essentially not meaningful enough for SLA. In the following, we will take a close look at these clusters to gain a better understanding of the research areas evolving over the years and provide tentative labels for them, as shown in Table 1.  This cluster is the largest and it contains 19 works, mostly published around 1994. The silhouette value of the cluster is 0.917, indicating a high consistency among the 19 cited works in this cluster. Table 2 includes the node information within Cluster #0. Each node represents one cited reference. The node information for each cluster including Cluster #0 is extracted from Appendix C. This study will only briefly look at 3 works with the highest citation frequencies in the cluster.  This cluster is the largest and it contains 19 works, mostly published around 1994. The silhouette value of the cluster is 0.917, indicating a high consistency among the 19 cited works in this cluster. Table 2 includes the node information within Cluster #0. Each node represents one cited reference. The node information for each cluster including Cluster #0 is extracted from Appendix C. This study will only briefly look at 3 works with the highest citation frequencies in the cluster. Ellis's book, The Study of Second Language Acquisition [9], is referenced most frequently (42 times) by all the articles in our dataset. A general framework for SLA study is put forward in this book that covers essential SLA topics such as learner language, the linguistic environment, the learner's internal mechanisms, individual differences in language learning, and research on classroom second language acquisition. The second most-cited work in this cluster is Pica [50], an article published in the Language learning journal and has been cited 25 times. This review article looks at the development of SLA in the 1980s with a focus on the social interaction and negotiation of L2 learners with reference to SLA classrooms. It argues that classroom negotiation facilitates learners' comprehension and structural segmentation of L2 input which contributes to their learning outcome. The third most-cited is Swain and Lapkin [51], which is focused on the mental process that leads to modified learner output. It argues that noticing a problem can 'push' learners to modify their output. Other highly cited works contained in Cluster #0, such as Donato [52] and Long and Robinson [53], are all related to learner's cognitive process and Vygotsky's Zone of Proximal Development sociocultural theory [15]. Given that the rest of the works in this cluster cover classroom interaction [52], student attention [54], and the input process [55], a more appropriate label for Cluster #0 appears to be Cognitive SLA Studies About Second Language Classroom.

Cluster #1
As shown in Table 1, the 2nd largest cluster (#1) has 14 members and a silhouette value of 0.936. The mean publishing year of the 14 works is 1987, and the detailed information of the works is listed in Table 3 with the top 3 most frequently cited works highlighted. Chaudron [56] reviews work on classroom-centered research and explains why student and teacher behaviors affect language learning with methodological discussions. Ellis [57] reviews classroom learning research and proposes a theory of instructed second language acquisition which has significant implications for language teaching. Taking a social and psycholinguistic perspective, Larsen-Freeman and Long [49] look at the empirical findings related to SLA by children and adults with an emphasis on how to design and carry out research in SLA. As this cluster also contains some important theoretical works such as Swain's Comprehensible Output Hypothesis [27] and Chomsky's and White's discussion of language origins [16,58], Theoretical SLA Studies and Their Applications appears to be a more appropriate label.

Cluster #2
As shown in Table 1, the 3rd largest cluster (#2) has 14 members and a silhouette value of 0.943. The mean publishing year of the 14 works is 2002, and the detailed information of the works is listed in Table 4 with the top 3 most frequently cited works highlighted. Schmidt [59] claims that attention or attended learning is far superior to other forms of learning and it holds the key to understand SLA and the development of interlanguage. Ellis's Task-Based Language Learning (TBLL) [60] is the state-of-the-art work in language learning and teaching, and it is no doubt one of the most influential theories in SLA that postulates the relationship between task and SLA, social aspects of TBLL, course design, teaching methodology, and the evaluation process. Norris and Ortega [61] carry out a meta-analysis to summarize both experimental and quasi-experimental studies into SLA instruction between 1980 and 1998. The other highly cited works in this cluster provide critical reviews of the research into Input Hypothesis [62,63], Output Hypothesis [64], and the social contexts of SLA teaching and learning [65]. In summary, works in this cluster are related with SLA teaching, learning, and research, and it is more appropriate to be labeled as Methods of SLA Teaching, Learning, and Research.

Cluster #3
As shown in Table 1, the 4th largest cluster (#3) has 12 members and a silhouette value of 0.916. The mean publishing year of the 12 works is 1988, and other detailed information of the works is listed in Table 5 with the top 3 most frequently cited works highlighted.  [67] divide learning strategies into metacognitive strategies, cognitive strategies, and social-affective strategies. Gardner [68] mainly discusses the social psychological factors (especially attitude and motivation). He distinguishes integrative motivation from instrumental motivation. The rest of the works in this cluster look at various learner individual strategies that affect second language acquisition. Based on the above analysis, we propose that this cluster represents SLA Studies of Learner Strategies.

Cluster # 4
As shown in Table 1, the 5th largest cluster (#4) has 11 members and a silhouette value of 0.96. The mean publishing year of the 11 works is 2009, and other detailed information of the works is listed in Table 6 with the top 3 most frequently cited works highlighted. Works in this cluster are mainly journal articles that report on the results of empirical studies. Based on a large scale study of Spanish/Swedish bilingual adults, Abrahamsson and Hyltenstam [69] argue that it is unlikely that adult learners of L2 will attain native-like fluency.
Lyster and Saito [70] reported on the effectiveness of oral corrective feedback (CF) based on a meta-analysis of 827 students. Li [71] reported on the effectiveness of corrective feedback in SLA based on a meta-analysis of 22 published studies and 11 Ph.D. dissertations. Both studies found that CF had a significant effect on L2 learners. In addition, works in this cluster also address the effects of grammatical complexity on learner language accuracy and fluency [72,73]. Therefore, we propose Teaching Approaches to Complex Tasks as the label for Cluster #4.

Cluster #5
As shown in Table 1, the 6th largest cluster (#5) has 10 members and a silhouette value of 0.947. The mean publishing year of the 10 works is 2007, and other detailed information of the works is listed in Table 7 with the top 3 most frequently cited works highlighted. Dörnyei's two works [74,75] review learner variables that affect SLA learning from the perspectives of other fields, such as psychology and sociology. The factors include personality, temperament, emotion, ability, motivation, style, strategy, anxiety, creativity, willingness to communicate, self-esteem and conviction. Ellis's book [76] takes a variety of perspectives, including cognitive, linguistic, sociocultural, and neurolinguistic to look at SLA development since the 1960s, and its coverage includes learner language, linguistic environment and social context, internal mechanisms, learner differences, and SLA instruction. The rest of the works in this cluster predominantly focus on the psychology of the language learner in second-language learning [77][78][79][80]. Therefore, we propose that this cluster represents Psychology studies in SLA.

Cluster #6
As shown in Table 1, the 7th largest cluster (#6) has 5 members and a silhouette value of 0.973. The mean publishing year of the 5 works is 2003, and the detailed information of the works is listed in Table 8 with the top 3 most frequently cited works highlighted. In this cluster, White's work [58] is a textbook for researchers that provides an overview and analysis of current SLA research conducted within the Chomsky's [17] generative linguistic framework and UG. She argues that SLA is constrained by principles and parameters. Using experimental psycholinguistic techniques, Clahsen [81] investigates grammatical processing in language learners and explains how grammatical processing in language learners differs from that of native speakers. Hyltenstam and Abrahamsson's The handbook of second language acquisition [82] discusses maturational constraints on SLA based on the some of the most important SLA hypotheses such as Critical Period Hypothesis, the Completeness Hypothesis, the Competition Hypothesis, the Fundamental Difference Hypothesis and the Input Hypothesis. It appears that none of the cluster labels discrete stage or evidence suggested by LLR or LSI algorithm is appropriate. Therefore, we propose Linguistic and psycholinguistic approaches as the label for this cluster.

Cluster #7
As shown in Table 1, The 8th largest cluster (#7) has 3 members and a silhouette value of 1. The mean publishing year of the 3 works is 2002, and the detailed information of the works is listed in Table 9. There are only 3 works in this cluster, and the most frequently cited work is N. Ellis's article [83] that takes an applied linguistic approach to discuss the input frequency of phonology, phonotactics, reading, spelling, lexis, morphosyntax, formulaic language, language comprehension, grammaticality, sentence production, and syntax. Tomasello [84] is focused on the usage-based theory of child language acquisition that develops from cognitive science, linguistics, and developmental psychology. It provides a competing theory against Chomskian generative grammar. Wray [85] develops a new computer-supported language processing system, TALK, to promote conversational fluency in non-speaking individuals. The cluster label Usage-based approaches suggested by LLR algorithm appears quite appropriate.

Cluster #8
As shown in Table 1, the 9th largest cluster (#8) has 3 members and a silhouette value of 0.925. The mean publishing year of the 3 works is 1993, and the detailed information of the works is listed in Table 10.

ClusterID
Author Year Source Freq. 8 Omaggio HA 1993 Teaching language in context 17 8 Lee J 1995 Making communicative language teaching happen 18 8 Horwitz EK 1991 Language anxiety 15 Lee and VanPatten's book [86] is basically a textbook for graduate teaching assistants and undergraduate teacher education majors. It provides a guideline for developing classroom environments to accommodate various listening, speaking, reading, and writing instructions. Omaggio [87] talks about teaching language in context. It is also a reference book for language instructors. Horwitz and Young's book [88] covers a number of areas of language acquisition, among which language anxiety emerges as one of the most important issues that institutions have to respond to. Therefore, the label Collaborative teaching suggested by the LLR algorithm represents the content of this cluster.

Cluster #9
As shown in Table 1, the 10th largest cluster (#9) has 3 members and a silhouette value of 1. The mean publishing year of the 3 works is 2009, and the detailed information of the works is listed in Table 11. The cluster label of reading comprehension for this cluster by LLR algorithm is acceptable since reading is the common topic in the 3 works contained within this cluster. However, a label of Reading and vocabulary is recommended after a close examination of the 3 works. Grabe [89] provides a critical discussion on a wide range of reading theories, such as implicit and explicit learning, attention, noticing, consciousness, and contextual and background knowledge in reading comprehension. Schmitt's book [90] provides an overview of a wide range of vocabulary research methodologies including discussions on validity and reliability of research on vocabulary as well as resources available such as lexical test papers, corpora, software, and testing websites. Schmitt [91] overviews current research on vocabulary learning and proposes word families for reading and oral discourse. Obviously, CiteSpace's LLR algorithm has not captured vocabulary when generating the cluster label.

Cluster #10
As shown in Table 1, the 11th largest cluster (#10) also has 3 members and a silhouette value of 1. The mean publishing year of the 3 works is 1994, and other detailed information of the works is listed in Table 12. Dörnyei [92] maintained that motivation is the main determinant of L2 learning achievement. This study investigates the nature and role of motivation in the L2 learning through an experiment. Oxford and Shearin [93] suggested several ways to extend learning motivation based on Gardner's [68] motivational theories. Applying social-psychological constructs, Clément, Dörnyei, and Noels [94] found that L2 student attitude, anxiety, and motivation toward learning English are associated with achievement. As all these works are related with motivation, the cluster is therefore labeled as Motivation in SLA.
To better visualize the origin, historical development, and current status of SLA knowledge clusters, we also performed a visual timeline analysis on the co-citation clusters of cited references as shown in Figure 4.
The map in Figure 4 is a little different from that located in Figure 3. This map shows the time span of each cluster, which bears some important research implications. It is obvious that Clusters #4, #5, and #9 are research areas that have emerged most recently. Particularly, we could expect that research topics within Clusters #4 and #5 will continue to receive attention for some time in the future. Figure 4 also shows that the research priority in SLA keeps changing over time. In the earlier part of 1987-2003 (Clusters #1 and #3), researchers were mainly interested in theoretical SLA studies and their applications as well as learner strategies in SLA; then some researchers changed to focus on cognitive SLA studies about classroom teaching, collaborative teaching, and motivation in SLA (Clusters #0, #8 and #10). From 2003 to 2010, the research interests in SLA shifted to working methods of SLA teaching and learning, linguistic and psycholinguistic approaches, and usage-based approaches (Clusters #2, #6, and #7). In most recent years, teaching approaches to complex tasks, psychology studies in SLA, and reading and vocabulary (Clusters #4, #5, and #9) were the most attractive research themes.
The nodes in the map are important indicators. Each node represents one cited reference, which is depicted with a series of tree-rings across multiple time slices. The area of each node is proportional to the total co-citation frequency of the associated reference. Table 13 lists the top 10 most frequently cited references. Except for Lantolf [95], all of the most frequently cited references fall into Clusters #4 and #5. This indicates that Clusters #4 and #5 are the most active research areas containing works that are cited most frequently. The nodes in the map are important indicators. Each node represents one cited reference, which is depicted with a series of tree-rings across multiple time slices. The area of each node is proportional to the total co-citation frequency of the associated reference. Table 13 lists the top 10 most frequently cited references. Except for Lantolf [95], all of the most frequently cited references fall into Clusters #4 and #5. This indicates that Clusters #4 and #5 are the most active research areas containing works that are cited most frequently.  So far, we have examined the largest 11 clusters of the co-citation analysis, which represent the 11 major research areas in SLA between 1987 and 2018. Since all articles included in the database are retrieved from top-tier journals indexed by WoS databases, the quality of their references is expected to be sound. As a result, the information obtained from this analysis will be reliable. For example, the  So far, we have examined the largest 11 clusters of the co-citation analysis, which represent the 11 major research areas in SLA between 1987 and 2018. Since all articles included in the database are retrieved from top-tier journals indexed by WoS databases, the quality of their references is expected to be sound. As a result, the information obtained from this analysis will be reliable. For example, the citation frequencies of key references reflect their relative importance and contributions to SLA research. The identification of these articles, therefore, can guide researchers, particularly Ph.D. students, to quickly grasp the most valuable works available in the field.

Pivotal Works in SLA
To identify pivotal literature in a research field is crucial for a researcher to understand the theoretical underpinning of a new research article and its connections with other fields of study with which the research may not be familiar. Pivotal literature is measured by the Centrality values of references produced along with the network map during a co-citation analysis in CiteSpace. Centrality values are normalized to the unit interval of [0, 1]. A node of high centrality is usually one that connects two or more large groups of nodes. In general, a reference with a Centrality value equal to or greater than 0.1 can be considered a key study [47]. Key studies are shown in a cluster network as a purple circle. A thicker circle indicates a higher Centrality value. Table 14 specifies the top 10 key articles in the literature in the SLA knowledge map. The top-ranked items by Centrality values are Dörnyei [75] in Cluster #5, with a Centrality of 0.03 and Pica [96] in Cluster #1, with a Centrality of 0.03. The Centrality values of the remaining 8 references are all 0.02. As a result, there are no purple circles in Figure 4. The results indicate that there may be a lack of pivotal studies in SLA. Although Chen's specification of a Centrality value of 0.1 as the threshold for key studies is based on his experience of analyzing data from biological or medicine fields [47], this value has worked well for studies using data from various disciplines [97][98][99]. The dataset of this study is even larger than some studies that produce pivotal works successfully. Therefore, the low Centrality values for all works in this study are not due to our data size. It more likely reflects the true situation facing SLA studies, and it echoes Gass's concern that to become a discipline integrated into the academic community SLA studies must be "needed" by other disciplines [48]. In fact, SLA is relatively a specific and narrow research field compared with some broader research fields. Therefore, it is quite natural for SLA researchers to cite and apply theories and practices from those general fields rather than the other way around. This may explain the lack of pivotal works with high Centrality values in this study.

Emerging Trends in SLA
In addition to generating network maps, CiteSpace can also reveal the frontiers and research trends in different periods. It uses an algorithm called Burst Detection to detect sudden growths in research interest to a particular work when its citation count increases abruptly. Compared with general frequency counts related to other works, the burst detection in CiteSpace can identify research frontiers more accurately [45]. Table 15 shows a ranked list of articles with their highest citation bursts taken together with their impact durations as based on the networks in Figure 3. The impact durations are indicated by blue and red lines in the column titled "1987-2018". The duration in which no bursts have been detected is depicted as a blue line, and the period during which a citation burst (CB) has been detected is shown as a red line. The start and end of a CB are also provided under the "Begin" and "End" columns. The table only shows 5 works that are experiencing sudden growth of citations towards 2018; works with CBs that have already ended can be found in Appendix A. Among the works having CB detected, Abrahamsson and Hyltenstam's paper [69], titled Age of Onset and Nativelikeness in a Second Language: Listener Perception Versus Linguistic Scrutiny, has the strongest CB strength (17.6663), followed by Norris and Ortega's paper [72], titled Towards an Organic Approach to Investigating CAF in Instructed SLA: The Case of Complexity (15.2459). "Ultimate attainment in L2 inflection: Performance similarities between non-native and native speakers" [100] ranks 3rd for CB strength (13.2445), while García's Bilingual and Multilingual Education in the 21st Century [101] ranks 4th (12.2491), followed by Paradis [102] ranking 5th (10.4661).
Contrary to Abrahamsson and Hyltenstam's argument that it is unlikely for adult L2 learners to attain native-like fluency [69], Hopp reports four experiments to test whether L2 inflection problems in L2 learners persist or they are a consequence of age-related grammatical impairment of L2 morphosyntax [100]. The study finds that native-like and ultimate attainment of L2 inflection is possible. It appears that nativelikeness is still the ultimate goal for many learners, which attracts continuous exploration from SLA researchers. The measurement of complexity, accuracy, and fluency (CAF) covered in Norris and Ortega's paper [72] is related to automatic oral language generation. Its high CB strength indicates that a growing number of researchers are becoming optimistic about the future of automatic language generation, probably as a result of the fast development of computers, the Internet, and artificial intelligence (AI) technology. García's book [101] addresses a number of global and local concerns in the 21st century such as bilingualism, bilingual education (both theories and practices), and international bilingual education policies with a focus on US and EU language policies in education. Paradis [102] provides critical views on neurolinguistic aspects of bilingualism. It examines the roles played by declarative and procedural memory to the L2 processing mechanism. Its high CB strength indicates that the application of neurolinguistics in SLA is receiving increasing attention. This is also a developmental direction for SLA that has been expected for some time [5,8].
Based on the outcomes above, we have identified the current research frontiers in SLA as bilingualism studies, automatic language generation including AI and computer-aided learning, language policies regarding SLA, and neurolinguistics. Among these research frontiers, bilingualism studies and language policies appear more global than local SLA-specific issues as they are also found as the topics experiencing a significant increase of interest in the field of applied linguistics [41]. The finding of neurolinguistics being another research frontier in this study also supports Lei and Liu's [41] findings that interest in the topic of eye-tracking has risen significantly during 2013-2016 as eye-tracking is one of the most important neurolinguistic research methods. Automatic language generation had not been captured by previous studies as a research topic receiving significant attention [30,41]. However, its identification as a research frontier in this study is not surprising since AI has enjoyed a huge popularity in recent years, particularly after the media exposure of Google's Alpha Go. In addition, since the time slice of this study was set to 8 years (2011-2018), it is likely that its citation number only burst in more recent years. In sum, the identification of these frontiers is based on the CB strength, and continuous attention is still needed in order to watch whether these frontiers will actually develop into research areas in the future. It is worth noting that works not listed in Table 15 may also have the potential to become research frontiers in SLA. As rightly pointed out by an anonymous reviewer, it usually takes at least two years for a publication to receive a meaningful number of citations. Therefore, works with a high CB value during 2011-2018 would be different if similar studies were carried out in the future.
Between 1987 and 2018, a total of 34 strong CBs have been detected, as shown in Appendix A, with their cluster numbers shown in Appendix C. The top-ranked item by burst strength is Lantolf J [95] in Cluster #15 (35.29) followed by Nation ISP [103] in Cluster #14 (29.68), Schmidt R [59] in Cluster #2 (27.31), Ellis R [9] in Cluster #0 (23.67), Ellis N [83] in Cluster #7 (22.83), Tomasello M [84] in Cluster #7 (21.19), and Oxford R [66] in Cluster #3 (20.53). All these works have experienced stronger CBs than those listed in Table 15 have. Although the former is no longer seen as leading research frontiers with the end of their CBs, these important works are still frequently cited when conducting SLA research.

Important SLA Keywords
To further explore the key research topics in SLA during 1987 and 2018, co-occurring analyses were performed using keywords in our dataset utilizing CiteSpace. We set the time partition at 8 years and the top 25 most frequently cited articles were selected in each time period. Figure 5 shows the timeline view of co-cited keywords, which puts keywords in order from left to right according to their years of publication. The left-sided ones are "old" ones, and the right-sided ones are "young" more recent ones.
In the early years, SLA research has scrutinized general concepts such as English, children's language, instruction, foreign language, comprehension, target language, native language, etc. Instead, it focuses on more specific aspects such as discourse, perception, context, and knowledge. Research topics in more recent years focus on education, motivation, proficiency, and vocabulary. Keywords significant difference, control group, and experimental group, located near the bottom of the figure highlight the strong experimental tradition in SLA research. All the keywords projected in Figure 5 are yellow, green, or brown (a mixture of green and yellow). This means that major keywords have been used consistently over the past 30 years. In other words, keywords used together frequently 30 years ago are still being used together by researchers in recent works.
Just as CBs may be detected for an individual piece of research work, bursts of keyword counts can also be detected [45]. Table 16 lists the top 20 keywords from the dataset with the strongest bursts in SLA research between 1987 and 2018.   The keywords appearing in this list are invaluable since they have been strongly focused on by various academics and have been expected to continuously provide insight into the development of new research areas within SLA studies. Furthermore, keywords outperform theories in disseminating new ideas and concepts and are mandated by all journals. Providing consistent keywords will benefit the overall communication and education in SLA research.

Conclusions
In this study, a critical review of SLA literature since 1987 has been carried out in a way that differs from traditional literature reviews. Driven by bibliometric big data, a review at the disciplinary level becomes possible, and the results of this overarching review have helped to clarify some issues that have been under debate for a long time (i.e., whether SLA is an integrated discipline and, what can be done to facilitate the discipline's integration into academic community).
Eleven major research areas in the discipline have also been identified, and they are visually presented with a clear timeline of information that provides a more effective guide for the potential The keywords appearing in this list are invaluable since they have been strongly focused on by various academics and have been expected to continuously provide insight into the development of new research areas within SLA studies. Furthermore, keywords outperform theories in disseminating new ideas and concepts and are mandated by all journals. Providing consistent keywords will benefit the overall communication and education in SLA research.

Conclusions
In this study, a critical review of SLA literature since 1987 has been carried out in a way that differs from traditional literature reviews. Driven by bibliometric big data, a review at the disciplinary level becomes possible, and the results of this overarching review have helped to clarify some issues that have been under debate for a long time (i.e., whether SLA is an integrated discipline and, what can be done to facilitate the discipline's integration into academic community).
Eleven major research areas in the discipline have also been identified, and they are visually presented with a clear timeline of information that provides a more effective guide for the potential reader. The visualization technique and CiteSpace software used in this study produce high-quality clustering maps with clear structures and sound explanatory power. The labels of the research areas have subsequently been subjected to manual scrutiny and refinement to ensure the objectivity of the results. A pivotal analysis of the reference data has shown a serious shortage of pivotal studies in SLA, which may contribute to the difficulties facing SLA discipline in its integration into a broader academic research landscape. In addition, a citation burst analysis may reveal the contemporary research frontiers in SLA available to subsequent inspection.
Although this data-driven review has achieved its research objectives, it should be pointed out that, like any bibliometric studies, this study also suffers from some limitations. First, despite careful manual examination, there may be missing information in our dataset. This study used a keyword search to collect SLA publications in the WoS database. This may lead to some articles being left out if the author did not use those terms in the article's title, abstract or author-provided keywords. A better practice is to search all articles in all SLA related journals indexed in the WoS database, as done in a recent bibliometric study of applied linguistics [41]. Second, as found in a number of studies, automated search results are not error-free [104][105][106]. Although this study has refined the search results to "linguistics or education educational research or language linguistics" categories, some erroneous entries may still exist within the dataset. In addition, the fast update of WoS databases requires the researcher to synchronize the dataset regularly, and during the synchronizing process, some duplicates may be generated. Another limitation relates to the selection of a time span of the WoS dataset in the co-citation analysis. As pointed out by an anonymous reviewer, it would be methodologically more appropriate to only include bibliometric data from at least the two years prior to the study when conducting co-citation analysis. Despite the above-mentioned limitations, this study has provided a new perspective on the development of SLS. Future studies looking at longer timespans are also strongly encouraged.
Funding: This research received no external funding.

Acknowledgments:
The author is grateful for the two anonymous reviewers who provided invaluable suggestions on methodological issues and the presentation of this paper.

Conflicts of Interest:
The authors declare no conflict of interest. The founding sponsors had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, and in the decision to publish the results.