The participants are 509 Chinese students, whose grades range from P₄ (fourth graders of the primary schools) to S₃ (third graders of the senior high schools). These students are from two primary schools, two junior high schools, and two senior high schools. Every pair of schools have similar quality in education, the same English syllabus, the same English textbook, and the same class hours each week. So the learning environments are very similar and we can thus minimize the interference of irrelevant factors. The English teachers were consulted about the English proficiency of these grades. They are very sure that on the whole, the students’ English proficiency of a higher grade is better than that of a lower grade, though the English test paper for the higher grade is more difficult than that for the lower grade and the average score of the former is even a bit higher than that of the latter. Grades are believed to be one of the most valid indexes for language proficiency (Wolfe-Quintero et al., 1998; Lu, 2011). Besides, to eliminate the “outliers” of each grade, a simplified version of questionnaire based on LHQ 2.0 (Li et al., 2014) was employed, including questions about age, gender, education, years of English learning, after-school English learning courses, overseas experience, and self-evaluation of English level. Average-level students whose language proficiency were representative of their grades were then selected as participants according to the questionnaire and their English quiz scores. Hence, in this study, student grade, i.e., the years of English learning, was an acceptable proficiency indicator of English. Therefore, writings by students from P₄ to S₃ are used to observe whether there is an emergence of scale-free and small-world syntactic network during the process of second language learning. The information of each grade, including the age and years of English learning, is presented in Table 1.

Writing tasks for all participants were assigned by their own English teacher in classroom tests at roughly the end of the semester in December. Students were fully aware that their compositions would be used for research, but their personal information would not be revealed under any circumstances. Written informed consent was obtained from participants above the age of 16, and the parents of all participants under the age of 16. The study was approved by the Research Ethics Board of Zhejiang University.

Students were required to finish their compositions in a prescribed period of time without any form of aid. Word count was prescribed (after consultation with the teachers in relevant schools) clearly for each grade in writing instructions: 40 words for fourth graders, 50 for fifth graders, 60 for sixth graders, 100 for the junior high school students and 130 for the senior high school students. In order to arouse the writing interests of the primary school students and to encourage them to write as much as possible, their topics such as “My Family,” “Fruit,” “Planting Trees” and “My Weekend” are presented in the form of colorful pictures; whereas for the high school students, the topics are presented in English, like “My Weekend,” “An Embarrassing Thing.” The topics include description and narration, because these two genres are not that demanding, and students are supposed to master these two genres at a very early stage of learning. The different topics are evenly distributed among the primary school students, but the junior and senior high school students used the same topics, respectively. By prescribing similar topics and genres, we intend to minimize the effects of irrelevant factors.

All the compositions were sent to us by their teachers, and were inputted into the computer after screening off the invalid ones, which are either too short (not meeting the minimal word count requirement of each grade), incomplete, irrelevant, or inappropriate. Besides, to ensure similar network sizes, which may facilitate the following comparative study, and guarantee consistency of the samples, it was decided that for each grade, the size of sample should be about 5000 word tokens. Therefore, 509 compositions were randomly chosen as the source of data for the current research, with a total of 45503 word tokens. Table 2 shows the information of the language materials, including the topic, the number of sampled compositions, and the number of word tokens of each grade.

In order to conduct complex network analysis, we have to construct language networks in the first place. However complex a network is, its basic elements are simple (Liu, 2008): vertices (or nodes) and edges (Newman, 2003, 2010). The former represents the entities of a network while the latter, the relationships between these entities. In a language network, vertices are linguistic units, such as words, and edges are pairwise relations between these linguistic units (Liu, 2008).

Suitable network representation is the prerequisite for the valid network analysis (Butts, 2009). To investigate the incremental syntactic development of L2 learners, syntactic networks are constructed, where a vertex represents a word type and the edge, the syntactic relation between two words. Besides, language networks must be constructed on the basis of linguistic rationales in order to be of research significance for language studies (Liu, 2011). Constituency, which concerns part-whole structures, and dependency, which concerns the binary asymmetrical relations between words, are two principal methods for syntactic analysis. In this study, the syntactic network is constructed on the basis of dependency structure. First, dependency grammar is shown to be more suitable for the research of language acquisition for learner language involving syntactic mistakes (Jiang and Ouyang, 2017). Second, the elements of dependencies are perfectly suitable to construct a network (Liu, 2008). Third, dependency grammar has been effectively applied to the investigation into L1 syntactic networks (Corominas-Murtra et al., 2009). For a reliable comparison between L1 and L2, dependency grammar is employed in the current study to construct syntactic networks.

According to dependency grammar, words in a sentence are connected by syntactic dependency relations (Mel’čuk, 1988; Hudson, 2010). In each dependency, there is a governor — the head word, and a dependent — the word governed by, or depends on the head word, as shown in Figure 1. The asymmetrical relation between the two words is shown by the arrow, which points from the governor to the dependent, with a label on the arrow to indicate the dependency type, or the syntactic relation. In Figure 2, for example, The depends on student; has governs student and book; book governs a; and has is not governed by any word (i.e., has is the root of the sentence). From the perspective of network, the dependent and the governor in sentences are vertices, and the dependency relations between them are edges.

Dependency structures of sentences can be presented in forms of tables, as illustrated by Table 3, which gives the dependency structures of two English sentences (This is a book and This book is very interesting).

These dependency-annotated sentences can be converted into a syntactic network, as illustrated by Figure 3, which displays the syntactic network converted from Table 3. The vertices in the network are word types, and the directed edges among them indicate the asymmetric syntactic dependency relations. In Table 3, a and This depend on book which in turn depends on is. In Figure 3, hence, there are two edges pointing from the vertex book to a and This, and one edge pointing from the vertex is to the vertex book.

Syntactic treebanks are constructed through a syntactically tagging software Stanford Parser 3.6.0¹ (Marneffe and Manning, 2008), and then are manually checked to ensure annotation accuracy. Details about this tagging software and the procedures are available in Ouyang and Jiang (2018). The syntactic errors in the students’ compositions are not corrected, and are left as they are in the treebanks as well as syntactic networks. The basic elements involved in network analysis are vertices and edges between them. One column of dependents plus another column of governors in the treebank will suffice for network construction. Therefore, syntactic errors, can only be reflected by the lack of edges between vertices in networks. For example, a learner made a syntactic error that the main verb is is missing in the sentence This is a book. Then, there would be no governor for the subject This and the object book in the treebank, and there would be no edge between vertices is and This and no edge between vertices is and book in the corresponding syntactic network. Take the sentence I must go to home as another case. The word to is redundant in this sentence, so there would be no governor for this word in the treebank and no edge linking to this vertex in the network. They will thus affect the network parameters such as degree and clustering coefficient. In this way, syntactic network can reflect language learners’ actual performance.

All the 509 compositions from nine grades were inputted into computer in plain text format, and each one was labeled with a unique code indicating the student’s school, grade, number and topic. Following the annotation principles of dependency grammar, nine syntactic dependency treebanks were automatically constructed through Stanford Parser 3.6.0, and then were manually checked. These dependency treebanks, each corresponding to a grade, were then converted into dependency networks by Createpajek² (a network converting software, whose details are available in Appendix). So 9 syntactic networks are generated, each labeled according to the grade, such as network P₄, network J₁, network S₃, and the like.

Network analysis may produce a variety of quantitative measures (Newman, 2010; Liu, 2011) to characterize the macroscopic properties of language networks (Cong and Liu, 2014). In the present study, the differences in network properties among grades may reveal the cognitive mechanism underlying the process of language learning. In fact, network science has offered effective instrumentalities to capture complex properties of cognitive and behavioral processes. For example, cognitive scientists have adopted network theory to describe neural organization and cognitive processes (Baronchelli et al., 2013; Karuza et al., 2016), to explore human learning process (Karuza et al., 2017; Kahn et al., 2018), and to investigate language acquisition and learning (Sizemore et al., 2018). In this study, the macroscopic network properties like scale-freeness and small-worldness can be indicated by network parameters, such as degree, average degree, average path length, clustering coefficient. On the basis of these parameters, the emergence of network properties and its linguistic implications will be explored.

Table 4 shows the general information of these networks from nine grades, including the number of vertices, the number of word tokens, and the Type/Token ratio (TTR). In these networks, vertices are word types, and the total number of words is the numbers of word tokens. To investigate lexical diversity/vocabulary richness of texts of similar size, the Type/Token ratio (TTR) is widely used (Richards, 1987). It is found that students from higher grades often present more lexical diversity or vocabulary richness. In addition, the global network of each grade (including all the vertices/word types) is presented in Figure 4 (see the drawing process in Appendix). Dots in graphs are vertices (word types), and lines among vertices are edges, i.e., syntactic dependency relations. The numbers of both vertices and edges increase steadily with the growth of grade.

To examine whether there is the emergence of scale-free property, the degree distributions of these 9 networks are extracted and compared with the degree distributions of their corresponding random networks (Newman, 2003), which may help identify the scale-free property of a network. For the purpose of reducing the noise in the long tail, the present study opts for the cumulative degree distributions (in log-log scales) of the nine networks, as shown in Figure 5.

It can be seen in Figure 5 that all the determination coefficients R² of nine syntactic networks are above 0.9, indicating that the degrees in 9 networks all fit well the power-law distribution, displaying Zipfian-like distributions. The fitting formulas are presented in each graph, with the power exponent γ’ and the determination coefficient R² presented in Figure 6, which registers a rapid increase of γ’ during the first three years of language learning, and then a tendency to stagnate at around 1.38 during the rest 6 years. The degree distributions of nine corresponding random networks are well fitted to binomial distribution, which is the essential feature of random Erdös–Rényi networks, and won’t be presented here due to the space limit. The Zipf’s law-like distributions of degrees in nine syntactic networks and the binomial distributions in 9 random networks suggest that the syntactic networks of all 9 grades exhibit a scale-free property.

Two network parameters, average path length (L) and clustering coefficient (C), are used to identify the small-world property of networks. Table 5 displays the average path lengths (L) and the clustering coefficients (C) extracted from the 9 syntactic network and the corresponding random networks.

Compared with the random network counterpart, a small-world network presents smaller average path length and far greater clustering coefficient (Watts and Strogatz, 1998). As Table 5 shows, all the nine syntactic networks have far greater clustering coefficients and smaller average path length than the corresponding random networks. These results suggest that all the 9 syntactic networks display small-world property.

This section will be devoted to the English lexical and syntactic development of the primary school students, who have developed neither vocabulary nor syntax well enough to express ideas freely or construct sentences flexibly. Consequently, they had to rely on their mother tongue to finish the writing, and they were allowed (not encouraged) to occasionally use Chinese words. To faithfully reflect their actual use of English, Chinese words in the compositions of the primary schools students (P₄, P₅, and P₆) are included in the analyses in the treebanks and the networks only in this section. The number and the percentage of Chinese word tokens in each treebank are presented in Table 6. The syntactic dependency networks of the three primary school grades are displayed in Figure 7. In each graph of Figure 7, for better visualization, English words are symbolized by vertices on the left circle, and Chinese words, by vertices on the right circle. The lines between the left and the right circles are syntactic dependency relations between English words and Chinese words.

It can be seen in Table 6 and Figure 7 that in the primary school years, the use of Chinese words tends to decrease with the advancement of L2 learning. Chi-square test shows that the difference in percentage of Chinese words is statistically significant between P₄ and P₅ (21.19 vs. 7.19%, p = 0.008). Accordingly, the cross-language syntactic relations between English and Chinese words decreased with the grade growth, which are presented by the lines between the left and right circles in the graph. These results quantitatively reveal that the fifth graders (P₅) have made considerable progress in their use of English, manifesting the importance of the first year in English learning.

A network has some important parameters, such as the number of vertices, average degree < k >, clustering coefficient (C), and the average path length (L), which can characterize the global network properties. As mentioned above, the primary school pupils often use Chinese words in their compositions. But when calculating these network parameters, Chinese words are excluded and only English words in the compositions are included for following network analysis, because the focus of our research is on the syntactic development of English.

The parameters of nine syntactic networks and nine random networks are presented in Table 5, and how these parameters change with the growth of grade is shown in Figure 8. The non-parametric Kruskal–Wallis test was employed to investigate whether there are significant differences between adjacent grades in these parameters. Figure 8A shows that the number of vertices monotonously increases with the growth of grade, and that the number of edges experiences a rapid growth when students move from J₁ to S₁. The changes in average degree are presented in Figure 8B, which shows that the average degree on the whole increases across the primary school and the junior high school years, except a slight decrease in J₁, but fluctuates violently in the senior high school grades. Kruskal–Wallis test suggests that there are no significant differences in degree distribution between adjacent grades. Table 7 displays the 5 vertices (words) with the highest degrees in each network. It can be seen that the nine syntactic networks have similar high-degree words that function as hubs: is/are/was, and, the, a, I, to, at, and in. Among them, is/are/was has the highest degree (the most edges) in 6 networks, especially the networks of the primary school grades. For S₂ and S₃ learners, the preposition to has the highest degrees and becomes the most important hub of the network.

The changes of clustering coefficient (C) are displayed in Figure 8C. It can be seen that the clustering coefficient descends dramatically during the first 3 years, rises slightly in the junior high school years, and then fluctuates in the senior high school years. Kruskal–Wallis test suggests significant differences between the grades of P₄ and P₅ (p < 0.001), P₆ and J₁ (p < 0.001), J₁ and J₂ (p = 0.040), J₃ and S₁ (p = 0.003). Figure 8D shows the changes in average path length (L), which experiences a sharp increase in J₁. Significant differences are found between the grades of P₆ and J₁ (p < 0.001), J₃ and S₁ (p = 0.010).

The language acquired by L2 learners, i.e., the “interlanguage,” is considered as closely related to both the native and the target languages (Selinker, 1969). Then it is valuable to compare network properties (parameters) of the interlanguage, the native language, and the target language, so as to pin down the possible unique features of interlanguage. The network parameters of native language (Chinese and English), which are from Liu and Li (2010), are presented in Figure 8B–D, where the full line indicates English network parameters and the dotted line indicates Chinese network parameters. The results of the comparison will be discussed in Section “Comparisons of Network Parameters of the Nine Networks.”

Scale-free and small-world properties are two network indictors of the emergence of syntax. With the abrupt appearances of these two properties in L1 networks, Corominas-Murtra et al. (2009) identified the emergence of syntactic network in L1 acquisition. The present study also uses these two properties as network indicators, and the results show that the networks from nine grades (P₄ to J₃) all exhibit scale-free and small-world properties.

A scale-free network means that in this network a minority of vertices have extremely high degrees while a majority of vertices have relatively low degrees (Newman, 2003; Ferrer-i-Cancho et al., 2004). To recap, the degree of vertices follows a Zipfian-law distribution, as can be seen in Figure 6, which clearly points to the fact that the nine syntactic networks all exhibit the scale-free property. Obviously, this finding is different from the findings of L1 study by Corominas-Murtra et al. (2009), who reported that L1 acquisition experiences a phase shift from a non-scale-free structure to a scale-free network when children were about 2 years old. However, our findings show that there is no such phase shift in L2 learning.

But there are differences among nine networks in how well the Zipfian distribution fits the degrees of their vertices. These differences are reflected by the power exponent γ of degree distribution, which often ranges from 2 to 3 in a scale-free network (Barabasi and Albert, 1999). The γ’ of the corresponding cumulative degree distribution equals γ-1, so the value of γ’ is between 1 and 2. The more γ’ approximates to value “1,” the better Zipfian distribution fits the data. As can be seen in Figure 6, γ’ first rapidly increases during the first 3 years of language learning, then stagnates at around 1.38. In a syntactic network, the scale-free property reflects the different ability of vertices (words) to syntactically combine with other words (Cong and Liu, 2014). Therefore, the fact that the degree of a syntactic network fits Zipfian distribution better implies that the students are able to use these words more proficiently. One would assume that with the growth of age and grade, the syntactic network should produce exponent closer to value “1,” fitting Zipfian distribution better, which, however, is not confirmed in our study. This could be due to the large quantity of language mistakes made by the primary school students and the repeated use of similar sentence structures. Table 7 shows that in networks P₄, P₅, and P₆, the top hub is invariably the vertex is, whose degree is constantly over 200, almost twice as many as the degrees of the top hub in networks S₁, S₂, and S₃. In other words, networks P₄, P₅, and P₆ display a more Zipfian-like distribution. However, the high degree of is in networks P₄, P₅, and P₆ does not mean high language proficiency. Instead, it may result from the beginners’ limited lexicon and the repeated use of similar sentence structures such as This is… and It is…. A detailed explanation would be given in Section “Comparisons of Network Parameters of the Nine Networks.”

Compared with the random network counterpart, a small-world network has smaller average path length and far greater clustering coefficient (Watts and Strogatz, 1998). As Table 5 shows, all the nine syntactic networks display far greater clustering coefficients and smaller average path length than those of the corresponding random networks. These results suggest that all the syntactic networks display a small-world property.

Similarly, there is no abrupt emergence of small-world property during L2 learning. In other words, the syntactic networks of L2 learners exhibit small-world property even in the primary school years. In a nutshell, the dependency networks of L2 learners from all 9 grades display both scale-free and small-world properties. In contrast, such properties, which are absent at the initial stage, will suddenly emerge during L1 acquisition. This disparity can be attributed to the fact that during L2 learning, learners can rely on the syntactic similarity between the native language—Chinese and the target language—English in this case to construct sentences. It is known that both languages are SVO languages, with some slight differences, such as the placement of attributives, adverbials and the word order of interrogatives. For language beginners, most of their productions are simple sentences with basic subject-verb-object structure like My family has six people. I have a brother. He is tall and The apple is haochi (Chinese pinyin for delicious). These simple sentences have similar structures in both English and Chinese. Therefore, as long as the basic vocabulary is mastered, early English learners can readily construct English sentences by resorting to their syntactic knowledge of Chinese. Moreover, English dependency network has much in common with Chinese dependency network (Liu and Li, 2010). In other words, positive transferring of mother language’s syntax is possible, which may explain the absence of sudden emergence of syntactic complex networks in Chinese English learners. The syntactic knowledge of Chinese facilitates the mastering of English syntactic structures. Therefore, if native and target languages have similar syntactic structures, syntactic phase transition may not take place during L2 learning. This tentative conclusion, of course, needs more evidence from different languages. In addition, this lack of phase shift can be explained with the Unified Competition Model (UCM) proposed by MacWhinney (2012). There is a competition between native language and target language during L2 learning: the “entrenched” L1 knowledge constrains the learning of new knowledge of L2, and the degree of “entrenchment” depends on the extent of consolidation of the learner’s L1 (Li, 2015). As L1 knowledge becomes more solid, new and different L2 structures are more difficult to acquire. Chinese English learners begin to learn L2 at around the age of nine, and at that time their L1 syntax has already been well-established. Hence, learners will use the already well-established L1 syntactic rules to generate L2 sentences. L2 syntactic network, as a result, bears much resemblance to extant L1 network, which renders the phase shift very unlikely in L2 learning.

There are three reasons for the extra attention paid to the primary school students. First, as mentioned in Section “Background Information of Participants and the Corpus Construction,” Chinese students start to learn L2 at the fourth grade in the primary school, at the age of nine. L2 research has attached much importance to children’s second language learning (Lakshmanan, 1995). Therefore, this period is worthy of close examination: if there is the sudden emergence of syntax, which is signaled by the small-world and the scale-free properties, it may well occur in this period. Moreover, language immaturity and the resulting vast mistakes pose a great challenge for studies of primary school students’ English. Previous studies rarely investigate into the second language of children in this period, not to mention the syntactic development. Not least, different from students of high schools, the primary school students often use L1 in their compositions. Given these reasons, this section will closely examine the language progress in the primary school students.

For the primary school students, neither the L2 vocabulary nor the L2 syntax has developed well enough for them to express ideas or construct sentences freely in L2. Consequently, they have to rely on their L1 to finish the writing (Cohen and Brooks-Carson, 2001). This code-mixing phenomenon is the most prominent characteristic of L2 learning in this phase. It can be seen in Table 6 and Figure 7 that for the primary school students, the use of Chinese words tends to decrease with the advancement of L2 learning, as has been reported in previous studies (Wang and Wen, 2002; Woodall, 2002). Besides, the dramatic decrease of Chinese words at the grade of P₅ is also noteworthy. As shown in Figure 7, the dependency relations between L1 words (the right circle) and L2 words (the left circle) decrease steadily with the grade growth. Most such dependency relations are replaced by English–English dependencies in networks P₅ and P₆. In brief, while L1 acquisition starts from scratch, L2 learning begins with the mixed use of two languages.

In order to explore the specific syntactic features exhibited at different stages of L2 learning, this section comparatively probes into the important network parameters of the 9 syntactic networks, including the number of vertices, the average degree < k >, clustering coefficient (C), and the average path length (L).

Table 4 shows the numbers of word types, tokens and TTRs of each network. It can be seen that the lexical diversity or vocabulary richness increases gradually with the grade growth. Figure 8A shows that the number of vertices increases steadily with the grade growth, and that the number of edges experiences, from J₁ to S₁, a rapid rise. Edges in a syntactic dependency network are syntactic relations between words. The rapid increase of edges thus indicates the rapid development of syntax. This means that the richness and complexity of syntactic relations, i.e., syntactic proficiency, have witnessed a substantial development in the junior high school year and the first senior high school year.

The changes in degree and average degree also deserve our attention. Vertices with relatively high degrees are hubs of a network. The 5 vertices with the highest degrees in the nine L2 networks are displayed in Table 7. L1 syntactic development witnesses a critical shift of hubs from semantic word it to function words a or the when the properties of scale-freeness and small-worldness emerge (Corominas-Murtra et al., 2009). However, in L2 syntactic development, the situation is different: all 9 syntactic networks have similar hubs (is/are/was, and, the, a, I, to, at, and in). Function words such as and, the, a have high degrees and play important roles even in the networks of L2 beginners. In addition, the copula be persistently has the highest degrees almost throughout the learning process. During the senior high school years, to has the highest degrees, becoming the most important vertex in networks. These findings are consistent with previous findings that function words such as articles and prepositions tend to be hubs of a network (Ferrer-i-Cancho and Solé, 2001; Cong and Liu, 2014; Jin and Liu, 2016). These hubs, i.e., the copula be and other function words, exists even in network P₄, which may account for the absence of the phase shift, that is, the sudden emergence of scale-free property in L2 learning. Function words are key elements in sentence structures (Klammer et al., 2010). The high degrees of function words indicate their great combinatorial capacity in syntactic networks.

The average degree, as shown in Figure 8B, presents an increasing trend over the period from P₄ to J₃, except for a slight decrease in J₁, and a descending trend in the senior high school years. The degree of a vertex (word) indicates its syntactic valency, i.e., its ability to form syntactic dependencies with words (Chen et al., 2011; Cong and Liu, 2014). Thus the said increasing average degree means enriched syntactic valency. That is, during the years of the primary school and the junior high school, students’ syntactic capacity develops gradually with the grade. However, during the senior high school years, the average degree firstly rises moderately, but then falls abruptly in S₃, which is beyond our expectation. Probably, for Chinese S₃ students, English learning has become a kind of repetitive practice and exercises for college entrance examination, resulting in the stagnation of syntactic development. As for the networks of mother tongues, the average degree is 7.127 for English and 6.478 for Chinese. Interestingly, the average degrees of networks J₂, J₃, S₁, and S₂ fluctuate within this range, somewhere between the average degrees of the native and the target languages. In other words, the interlanguage, namely the English acquired by Chinese students, is syntactically different from the target language. The disparity between interlanguage and target language is also revealed by the slight differences in the probability distribution of dependency distance (Ouyang and Jiang, 2018).

The changes in clustering coefficient are linguistically significant as well. As displayed in Figure 8C, the clustering coefficient descends dramatically during the first 3 years, rises slightly in the junior high school years, and then fluctuates in the senior high school years. The decrease of clustering coefficient in the first 3 years, especially in J₁, can be attributed to the rapid increase of vertices, i.e., vocabulary, and the resulting “dilution” of edges, which may be construed as the lagging development of syntax. Note that the clustering coefficient (C) of a vertex is the ratio of the actual edges to all possible edges among its neighbor vertices. Hence, if a vertex links to only one vertex (i.e., poor combinatorial capacity), the C of this vertex is zero. The sudden drop in J₁, therefore, implies that many vertices may have zero clustering coefficient, owing to the rapid increase of vocabulary and the simple syntactic structures used by beginners. In other words, J₁ students may experience an expansion of vocabulary, but lack the ability to use these words creatively in new contexts. But with the advancement of learning, this ability gradually improves, leading to enriched syntactic valency and increased clustering coefficient. Verspoor et al. (2012) proposed that the learning focus for beginners is the accumulation of vocabulary, and given that the vocabulary threshold is reached, middle-level learners will shift much attention to sentence structures. Besides, networks of native English and native Chinese display quite similar clustering coefficient: 0.122 and 0.128 respectively. The clustering coefficient of the interlanguage, except for the primary school students, generally fluctuates around that of native languages.

The average path lengths (L) of these syntactic networks are shown in Figure 8D. L measures the network distance and the separation degree between a pair of vertices. As can be seen in Figure 8D, the average path length experiences a sharp increase in J₁. Interestingly, the changing trajectory of average path length seems to inversely mirror that of clustering coefficient. The reason for the decrease of C, i.e., the rapid increase of vocabulary, may also account for the sudden rise of L in J₁. When all vertices are linked to one center node, the L is the smallest, and the small value of L in a network has much to do with hubs (Watts and Strogatz, 1998; Nishikawa et al., 2002). Therefore, in networks P₄, P₅, P₆, the relatively high degrees of the hubs, i.e., many dependency of words like is, lead to low L. The abrupt increase of L in J₁ may also indicate that the syntactic ability lags behind lexicon, which has also been observed in L1 learners under the age of two, i.e., before the phase shift (Corominas-Murtra et al., 2009). Moreover, the boom of edges during the junior high school years (as seen in Figure 8A) may have much to do with the reduction of the average path length L. As to networks of L1, the average path length is 3.308 for English and 3.371 for Chinese. It can be seen that, for both junior and senior high school students, the average path length is slightly below those of both native English and native Chinese. This fact reflects the independence of the interlanguage to some extent. Similar to average degree, the average path length of interlanguage does not reach that of target language. This fact suggests that, even for advanced Chinese English learners, the L2 syntactic network is sparser than the network of native English, which means that it is not as well-organized as the syntactic network of native English.

From the above analysis, it can be concluded that in L2 learning there are no syntactic phase shift, namely, no sudden emergence of scale-freeness and small-worldness in syntactic networks of Chinese English learners. Different from L1 acquisition, which starts from scratch and features the sudden emergences of the above two complex network properties, L2 learning is usually based on an existing L1 syntactic network. Therefore, such an emergence is absent. Instead, there is a rather gradual approximation to the target language. Moreover, foreign language is acquired non-linearly and different linguistic subsystems (lexicon, syntax, etc.) are learned at different paces in different learning stages. Vocabulary goes through an expansion in the first year of the junior high schools, because in that year English becomes a major subject and textbook provides considerable vocabulary input. Syntax, on the other hand, does not substantially develop until late junior high school years and senior high school years, especially J₃ and S₁. In other words, syntax develops later than lexicon. In addition, the fluctuations of some parameters in senior high school years support the theory of fossilization (Selinker and Lamendella, 1978). In other words, language competence of L2 learners does not improve continuously with grade. Instead, after a certain period, the progress may slow down, stagnate, or even decline. Besides, through the comparison of these network parameters of the interlanguage, the native language and the target language, it can be seen that interlanguage has its independent status, not simply a variety of the target language that is distorted by the mother tongue. It actually displays properties which are unique, peculiar to itself.