Quantitative Research on Chinese Sentences Structure Based on Pattern Grammar

Abstract

With the proliferation of corpora, various syntactic analysis methodologies have been developed. However, syntactic analysis of Chinese sentences demands a theory that focuses more on word order and the interaction between content and function words, which is satisfied by pattern grammar theory. This study investigates the effectiveness of pattern grammar theory in the analysis of Chinese sentence structure and explores the linguistic characteristics of official documents by establishing a description system for structural complexity in Chinese. Statistical methods such as skewness, kurtosis, and correlation analyses were employed to evaluate the system’s performance. The findings confirm the effectiveness of pattern grammar in analyzing Chinese sentence structure by showing that it can accurately reflect the unique properties of Chinese official documents, such as the preference for fixed collocations, declarative sentences, and flat structures. Based on the above findings, this study provides a semantically compatible and innovative theoretical system for analyzing Chinese sentence structures.

Keywords

pattern grammar sentence structure quantitative research

Introduction

Descriptions of sentence construction, which are closely related to the evolution of language and the universal cognition of human beings, also have great application value in the field of natural language processing, especially in the syntactic parsing of corpus-based studies. Therefore, developing an appropriate framework to describe sentence construction is critical. Such an urgently needed method is expected to describe different languages with distinctive features, including Chinese, which was the focus of this study.

Currently, the most widely adopted descriptions of sentence constructions are based on phrase structure grammar (PSG) (Chomsky, 1957) or dependency grammar (DG) (Tesnière, 1959).

PSG, which assumes that sentences are composed of phrases, is a generative device (Chomsky, 1957). According to the minimalist program, sentences are built by merging constituents in pairs to form larger phrases (Chametzky, 2003; Chomsky, 1995). Consequently, with many nonterminal nodes, sentence structure is represented by labeled tree diagrams. Some scholars have used PSG to analyze Chinese and investigate certain sentence structures, such as the structure of “de,” the Chinese appositive structure, and the syntactic structure related to personal pronouns (L. Liu, 2015; Yao, 2010; Y. Zhang, 2013; Zheng, 2013). PSG is also widely used in the automatic parsing of Chinese syntax, such as the Chinese Penn Treebank, the Stanford Parser, and the phrase structure parser developed by the Readworld Language Information Processing Institute.

DG, a modern concept advanced by Tesnière (1959), assumes that a sentence is constructed by dependency relations between individual words. This dependency constitutes a binary relationship between two linguistic units that is always asymmetrical because one is the governor and the other is the dependent (Hudson, 1990, 2007; Tesnière, 1959). Consequently, the syntactic analysis of a sentence is illustrated by an acyclic graph, in which arcs link the two units, showing their dependency relations. Many researchers have adopted dependency grammar to study noun phrases (Zhou & Huang, 1999), syntactic ambiguity structures (Yuan et al., 1999), and parallel structures (Zhao et al., 2009) in Chinese and constructed a multi-view Chinese treebank for sentence analysis (Qiu et al., 2015).

However, researchers have also noted deficiencies of PSG and DG in describing Chinese sentences. First, both PSG and DG are generalized from inherent grammatical capabilities; thus, to some degree, they are not real (Hunston & Francis, 2000). Because of the two distinctive characteristics of modern Chinese grammar: (1) word order is the foundation of Chinese grammatical structure (Lian, 1993; L. Wang, 2015; Wen & Hu, 1984); and (2) function words constitute an essential grammatical means in Chinese (Lian, 1993; Lu & Ma, 2003), and because PSG and DG are non-linear, the two prevailing systems fail to best represent Chinese sentence structure, where word order plays a significant role in grammar (Hunston & Francis, 2000). Moreover, the interactive relations between functional words and content words in Chinese sentences are not fully pictured and described by PSG and DG because the role of functional words is not emphasized in the two frameworks.

To address this problem, pattern grammar, an innovative analytical framework with a distinctive theoretical basis, has attracted increasing interest from scholars. Pattern grammar was chosen for this study mainly due to its four unique advantages, especially in Chinese language research.

Pattern grammar closely related to functional words. A pattern is a phraseology frequently associated with a word, particularly in terms of prepositions, groups, and clauses that follow the words. The coding of patterns, including actual functional words such as ba, de, le, and guo, and categories of content words such as V, N, and ADJ, can describe the governing position of function words over content words (see Figure 1). As a result, pattern grammar can be used to analyze the Chinese language from a new perspective, which depends more on functional words, word order, and context.

Pattern grammar emphasizes lexis and context. Theoretically, there is no distinction between lexis and grammar. Communicators express abstract concepts using lexis-suitable syntactic structures to meet their communication needs (Hunston & Francis, 2000). In pattern grammar, grammar refers to the logical consequences of looking at words in a context (Hunston & Francis, 2000). Therefore, it is suitable for describing context-dependent grammar in Chinese based on word order.

Pattern grammar describes real language in different authentic contexts based on a large corpus (Hunston & Francis, 2000), which is distinct from the traditional view that language is generated by inherent abilities. Instead of human-summarized rules, pattern grammar is based on empirical evidence and can better reflect natural language in the real world.

Pattern grammar involves a linear analysis of sentence construction rather than a hierarchical analysis (Brazil, 1995). It asserts that the process of generating a sentence can be regarded as the continuous fulfilment of new “prospections” (Sinclair, 1995), which complies with the psychological process of language use, especially in speaking. Therefore, pattern configurations can describe the complexity of a language theoretically (Hunston & Francis, 2000).

Although pattern grammar has much in common with construction grammar, as summarized in Hunston’s (2019) study, pattern grammar and construction grammar differ in the following aspects: First, construction grammar developed within the traditions of cognitive linguistics, rather than in the traditions of corpus linguistics, and until recently, there has been little dialogue between the two. Second, patterns in pattern grammar are purely observational phenomena, whereas constructions attempt to model the mental representations of language. Third, compared with patterns, constructions are studied at a high level of specificity, failing to progress towards a systematic description of a language, for example, the “affected for” construction, the “proxy for” construction, the “reactive for” construction, etc. The potential number of constructions is vast; their listing seems impossible, while pattern grammar does not presuppose a division into lexis and grammar (Sinclair, 1991, p. 3), which is more general and abstract, making it suitable for the grammatical analysis of sentences in running texts. For instance, the “proxy for” construction and the “reactive for” construction can be regarded as “N for” pattern and “ADJ for” pattern in pattern grammar. Fourth, patterns are grammatically linear with configurations of “pattern flow,”“pattern strings,” and “pattern loops,” which makes it an ideal grammar for the description of sentences in grammar research. “Pattern flows” occur when a word that is part of the pattern of another word has a pattern of its own (Hunston & Francis, 2000). “Pattern strings” represent a configuration when patterns do not overlap and a different configuration occurs. “Pattern loops” represent a configuration when it begins with one pattern, flows into another, and then loops back to the final element of the first pattern (Hunston & Francis, 2000), for example (see Figure 1).

Figure 1.

Examples of pattern flow, pattern string, and pattern loop.

As a remarkable complement to traditional grammatical theory, some researchers (Hunston & Su, 2019; Y. Liu & Lu, 2020; Perek & Patten, 2019) have adopted a pattern-driven approach to identify grammatical patterns, and scholars have begun to evaluate the performance of pattern grammar when it is transferred to a Chinese context. Y. Wang (2008) was the first to introduce pattern grammar in China. He systematically interpreted the theoretical origins, major propositions, and empirical applications in the Western context. X. Wang et al. (2018) analyzed the effectiveness of pattern grammar by illustrating the interaction between content and functional words and accurately reflecting the linear structure of run-on sentences. Zhan (2021) made a comparison between the concept of pattern and construction in Chinese context. And most pattern grammar researches on Chinese (Q. Fang, 2020; X. Wang et al., 2018; Xu & Qiao, 2017; D. Zhang & Qu, 2021) has been conducted at level of specific single patterns, instead of a sentence structural level.

Unfortunately, although a few investigators have analyzed Chinese from the perspective of pattern grammar structure, insufficient effort has been devoted to domesticating this theory in the Chinese context. No large-scale linguistic data analysis has been conducted to verify the fitness of pattern grammar for analyzing Chinese linguistic materials. No concrete statistical parameters have been reported; thus, a credible conclusion with robust quantitative support has not been obtained because no automatic software has yet been developed to describe Chinese sentence structures or to parse Chinese based on pattern grammar. In this context, establishing a Chinese sentence structure system from the perspective of pattern grammar is innovative and meaningful. Establishing this system will facilitate statistical research to describe Chinese linguistic characteristics through pattern grammar, with the purpose of validating the adaptability of pattern grammar in Chinese.

Given this context, the first aim of this research is to prove that pattern grammar, even if not superior to prevalent syntactic analysis theory, should at least be evaluated as an equivalently effective tool for the analysis of Chinese sentence structures. Specifically, this research aims to establish a framework and describe Chinese sentence structures based on the pattern grammar theory. The following hypothesis was proposed, along with a set of research questions.

Hypothesis: Pattern grammar can be treated as an effective theoretical framework for describing the structural characteristics of Chinese sentences, especially their structural complexity, as it can soundly reflect existing phenomena in formal written Chinese.

Question 1. How can we establish a systematic quantitative framework to describe Chinese sentence structures based on the pattern grammar theory?

Question 2. How are the pattern parameters distributed in the corpus, and how are they related to each other?

Question 3. Based on the distributions and relationships of the pattern parameters, what are the research findings of official Chinese sentence document registers from the perspective of pattern grammar?

The innovation of this research lies in two aspects: (1) adopting a new grammatical system to describe Chinese and (2) expanding the coverage of pattern grammar research on Chinese from the level of single patterns to sentence structures.

Methodology

Corpus

In this research, a self-established Chinese corpus is employed to study the structural characteristics of formally written Chinese. The corpus serves as a representative of the linguistic features, structure, and evolution tendencies of formal written Chinese, containing more than 1,860,000 words with texts concerning various aspects extracted from official Chinese documents from 1990 to 2018.

Prior to the regular procedures of pattern-marking, titles that are phrases, rather than sentences, are excluded to ensure the standardization of this corpus. The patterns of each sentence in this corpus were marked by a self-developed script named DepToPat with a high precision level of 94% in the pattern dimension and 88% in the sentence dimension. The program, which has developed user interfaces, segments linguistic data from certain registers and then transforms the segmented material from a perspective of dependency grammar, a dimension achieved by the Stanford Parser and the RNN (Recurrent Neuro Network) algorithm, to a perspective of pattern grammar through regulation expression wildcard matching. The annotation results of DepToPat present all possible patterns and their layers within a sentence in a Python list and Excel spreadsheet.

Based on the segmented text and annotated patterns, further configurations of pattern types, including pattern flows, pattern strings, and pattern loops, are distinguished by their original definitions.

Pattern Parameters

The five pattern parameters in Table 1 were designed and extracted based on Hunston and Francis’ (2000, p. 225) attempts to connect the distribution of pattern configurations with sentence complexity and the empirical research on Chinese sentence structure. WN, FN, and MPL reflect Hunston and Francis’s original theory, whereas LN and HPL are unique products of Chinese syntactic features.

Table 1.

Pattern Parameters.

Dimension	Parameter	Definition
General	WN	Number of words in a sentence
Depth	FN	Number of pattern flows
	LN	Number of pattern loops
	HPL	Number of pattern configurationlayers
Length	MPL	Mean length of patterns

WN displays the number of words in a sentence, referring to the sentence length. In this study, WN was measured using a word counter after tokenization. Moreover, because punctuation is generally not included within the framework of pattern grammar, sentence length is calculated according to the following equation:

\begin{matrix} WN = Tokenized Sentence Length \\ - Amount of Punctuation \end{matrix}

(1)

FN quantifies the number of pattern flows, defined as pattern configurations in which the first element of the latter pattern overlaps with the last element of the former pattern. Taking the sentence in Table 2 as an example, the pattern configurations [2]–[3] and [5]–[6] are regarded as pattern flows, which means that the FN of this sentence equals two.

Table 2.

Examples of Pattern Tokenization.

MPL weighs the average length of the patterns within a single sentence. This reflects the average number of elements in the pattern, and it can be calculated if the sum of pattern length (SPL) and the number of patterns (PN) are already known. In formula (2), $l_{j}$ represents the number of elements in the jth pattern.

Under these circumstances, the MPL of the sentences in Table 2 is 2.11.

SPL = \sum_{j = 1}^{PN} l_{j}

(2)

MPL = SPL / PN

(3)

LN portrays the number of pattern loops, a configuration in which the context begins with one pattern, flows into another, and then loops back to the element of the first pattern (Hunston & Francis, 2000). Although the pattern loop is not recommended as a benchmark of sentence complexity in English because of the differences between English and Chinese, it can still be considered an essential parameter reflecting the degree of complexity in Chinese sentences because Chinese is more inclined to propose complex attributes as an additional component of nominal noun phrases (Zhu, 1997). In addition, considering the frequent attributive nouns in formal written Chinese, pattern loops can explain the complexity caused by complex attributive nouns, just as the performance of pattern flows can be used in interpreting the structural function of English clauses and the “N…of…n” pattern. For instance, in the example in Table 2, pattern configurations [3]–[4], [6]–[7], and [7]–[8] are regarded as pattern loops that detect the complicated structure of modification ignored by pattern flows (LN = 3).

The HPL measures the number of pattern-configuration layers in a sentence. The pattern configuration layer is regarded as a side product of overlapped patterns whose relations contain all pattern flows, pattern loops, and a number of pattern strings. HPL is credible as a monitor of complex sentence structures because it embodies the influence of pattern flows and loops (HPL = 4).

All the above mentioned parameters describe Chinese sentences from a pattern grammar perspective. Among them, MPL suggests “length,” and FN, LN, and HPL jointly suggest “depth,” which contributes to the statistical analysis of pattern grammar through two dimensions. The descriptive system established using the five pattern parameters provides a practical answer to the first research question.

For further clarification, the pattern string, another element in pattern grammar theory, is not considered a proper parameter in this study. Because a pattern string is defined to describe adjacent patterns without overlapping parts, it is unlikely to cause any increase in the pattern layers or any fluctuation in structural complexity.

Data Analysis Procedures

To determine the performance of each pattern parameter to prove the self-consistency and effectiveness of pattern grammar and to describe the characteristics of Chinese sentence structures in official documents, this research follows the following procedures.

First, the distribution of pattern parameters is presented by reporting descriptive statistics for each parameter, skewness, and kurtosis coefficients of FN, LN, HPL, and MPL and by providing an overview of the processing complexity of the Chinese official document corpus. The probability density function (PDF) of each parameter is determined. Subsequently, the sampling skewness (SK) and kurtosis (KR) coefficients of each distribution are calculated to determine the inclination and concentration of language use. These coefficients are derived using the following equations, where x represents the value of FN, LN, MPL and HPL:

S K_{x} = \frac{\frac{1}{n} \sum_{i = 1}^{n} {(x_{i} - \bar{x})}^{3}}{{(\frac{1}{n} \sum_{i = 1}^{n} {(x_{i} - \bar{x})}^{2})}^{\frac{3}{2}}}, x = FN, LN, MPL, HPL

K R_{x} = \frac{\frac{1}{n} \sum_{i = 1}^{n} {(x_{i} - \bar{x})}^{4}}{{(\frac{1}{n} \sum_{i = 1}^{n} {(x_{i} - \bar{x})}^{2})}^{2}} - 3, x = FN, LN, MPL, HPL

Statistically, a positive SK indicates the majority of values less than the mean in the dataset, and a negative SK refers to the majority of values greater than the mean. When projected onto the distribution of pattern parameters, it suggests either a simple tendency or a more complex one in terms of the syntactic structure. Similarly, positive KR for each parameter reveals a more centralized performance of language processing difficulty, whereas negative KR indicates a more diversified situation. Generally, they represent preferences for certain language structures.

Second, a correlation between the pattern parameters and a linear regression analysis were conducted among the five pattern parameters to establish quantitative relations.

Third, this study aims to identify how WN influences other parameter distributions by delineating the probability density functions (PDFs) of FN, LN, MPL, and HPL under different levels of fixed WN. During this process, the contribution to the general skewness and kurtosis coefficients was examined from two perspectives: sentences with a fixed WN and sentences with a diverse WN.

Results and Discussion

In response to the three research questions, the results and discussion are divided into two parts: distributions of pattern parameters and relationships between pattern parameters. Specifically, in the second part, the research probes how sentence length influences the statistical distribution of the pattern flow, pattern loop, and pattern-configuration layers.

Distributions of Pattern Parameters

In Section “Distributions of Pattern Parameters,” the characteristics of Chinese sentence structures in official documents are discussed through descriptive statistics and distributions of pattern parameters, as well as their skewness and kurtosis.

Descriptive Statistics of Pattern Parameters

Among the five pattern parameters in this study, WN records the length of the sentences. FN, LN, and HPL indicate the complexity caused by the sentence “depth,” which represents the sum of continuation, dependence, and modification. MPL measures the complexity resulting from the structure within a pattern. Table 3 presents the descriptive statistics of the pattern parameters.

Table 3.

Descriptive Statistics of Pattern Parameters.

	WN	HPL	FN	LN	MPL
Mean	23.398	6.880	7.912	7.346	2.320
SD	15.158	2.623	5.009	4.686	0.300
Min	1.000	1.000	0.000	0.000	1.000
50% quantile	21.000	7.000	7.000	6.000	2.292
Max	105.000	23.000	39.000	40.000	7.500

As Table 3 shows, the range between the maximum and minimum values of each parameter is large, demonstrating a broad range of distribution in the Chinese official document corpus and confirming the validity of the corpus as a source for addition, more than 50% of MPL is located between 2.15 and 2.45, suggesting no significant variation, and the other four parameters exhibited similar characteristics. Moreover, the minimum of the pattern parameters indicates that the existence of sentences leads to the existence of patterns and pattern layers but not necessarily to the existence of pattern flows and pattern loops.

After the basic descriptive analysis, we depict the characteristics of the pattern parameters using their specific distribution, namely PDFs.

As shown in Figure 2, the PDFs of FN, LN, and WN exhibit asymmetric unimodal curves, while HPL displays a symmetric distribution. The distributions of FN and LN deviated significantly from normality. Therefore, instead of fitting the data with a normal distribution, asymmetric unimodal distributions, including chi-square and gamma distributions, were adopted to report the best-fitted distribution of these two parameters (Table 4).

Figure 2.

PDFs of FN, LN, WN, HPL, and MPL.

Table 4.

Parameters of FN and LN Distributions.

Parameter	Parameter	Coef.	SE	p > \|z\|	[0.025	0.975]
FN	Shape	2.012	0.060	.000	1.894	2.130
	Location	0.411	0.028	.000	0.356	0.466
	Scale	1.740	0.020	.000	1.700	1.779
LN	Shape	1.995	0.028	.000	1.940	2.049
	Location	0.521	0.024	.000	0.474	0.567
	Scale	3.422	0.040	.000	3.344	3.500

The distribution of FN achieved the highest similarity with a chi-square distribution, with robust statistical significance under the maximum likelihood estimation (MLE) criterion (p < .001). The black curve of FN in Figure 2 displays the PDF of the fitted chi-square distribution under very few conditions, such as FN = 6 and FN = 19, and deviates from the trend of the fitted curve. This deviation may be due to the quality of the corpus; a smoother PDF would be produced in a larger and more relevant corpus.

The distribution of LN displayed features markedly similar to the distribution of FN. Based on the MLE criterion, we determined a gamma distribution as the best-fit model for LN in Chinese official documents. Compared to the FN distribution, a flatter peak causes a transition from a chi-square distribution to a gamma distribution. Finally, we report the significance of the substitute, that is, a simpler gamma distribution with scale parameters (see Table 5), as a reference for other studies on similar registers.

Table 5.

Hypothesis Testing for LN Parameters.

Parameter	Coef.	SE	p > \|z\|	[0.025	0.975]
Shape	1.995	0.028	.847	1.940	2.049

Note. We conducted a t-test on the shape parameter.

Such an adjustment using an integer parameter is closer to the meaning of the gamma distribution in statistics. LN’s adaptation in fitting by frequently employed distributions indicates its efficiency. In other words, sentences in Chinese official documents involve many complex prepositive attributes that generate most of the pattern loops. To this extent, the phenomenon definitively proves the discovery of Lv (1982) and Li (2019), that is, the preference of Chinese official documents for complex prepositive attributes, such as “Shi-construction” and “Wei-construction.”

As Figure 2 indicates, the distribution of MPL is evaluated as a random variable with an irregular distribution. Most MPLs were between two and three, demonstrating that Chinese official documents tend to employ patterns of similar complexity within a sentence. Additionally, the truncation between values one and two suggests that, in formal written Chinese, the oral expression aspect is rejected and confirms the tradition of formal written Chinese identified by Yue (2005) and Feng (2018).

Skewness and Kurtosis of Pattern Parameters

Skewness and kurtosis coefficients can accurately reflect the preference for language structural complexity in certain language materials. As Table 6 shows, the skewness coefficients of all pattern parameters are significantly positive, implying that all their distributions are positively skewed with a long tail on the right; that is, the pattern parameters in most sentences are less than the mean.

Table 6.

Skewness and Kurtosis of Pattern Parameters.

Parameter	LN	FN	HPL	MPL	WN
Skewness	1.256	1.253	0.479	0.883	1.161
Kurtosis	2.201	2.131	0.871	10.518	1.693

As Table 6 shows, the skewness coefficients of all pattern parameters are significantly positive, implying that all their distributions are positively skewed with a long tail on the right; that is, the pattern parameters in most sentences are less than the mean. This phenomenon extends the asymmetric PDFs of pattern parameters to constitute evidence of a universal linguistic rule, the economy principle of language, and Zipf’s law, which expresses more information with fewer words, easier structures, and less effort (H. Liu, 2022; Zipf, 1949). Specifically, the skewness coefficients demonstrate that most sentences in Chinese official documents prefer concise structures with fewer LNs, FNs, and HPLs, and shorter MPLs. The realization of the economy principle of language verifies G. Fang’s (2001) effort to research the preferences of Chinese official documents for accurate and plain expressions.

The significantly positive kurtosis coefficients indicate a more concentrated peak in the distributions of each parameter, which indicates that most sentences in the Chinese official document show a strong tendency to adopt sentence structures with certain modes because they frequently present LN, FN, and HPL of a certain amount and MPL of a certain length. Consequently, the statistics verify the success-breeds-success principle of language, which asserts that the frequent usage of a certain language structure will further augment the probability of another use and predict the concentration and clustering of similar language structures (Price, 1976). This highlights the tendency of Chinese official documents to use specific sets of collocations and phrases with predetermined structures. This aligns closely with Yue’s (2005) finding that Chinese official documents, as statements on behalf of the agency, tend to employ relatively fixed collocations and expressions.

The absolute value of the kurtosis coefficient measures the degree of concentration and clustering of each parameter, thereby generating an evaluation of the fitted distributions mentioned in Section “Distributions of Pattern Parameters.” First, the kurtosis coefficient of FN and LN were 2.131 and 2.201 respectively. They are relatively close, and both are significantly positive, corresponding to the characteristics of a gamma distribution (or a chi-square distribution). Second, the kurtosis coefficient of the HPL was close to zero, suggesting its similarity to a normal distribution. Third, the kurtosis coefficient of the MPL reached a remarkably high level of 10.518, which demonstrates its validity in identifying clustering and concentration phenomena in the structures of Chinese official documents.

Relationships Between Pattern Parameters

In addition to portraying the distribution of single parameters, we aimed to identify the relationships among the different parameters. The overall correlation between the pattern parameters and the effect of sentence length on the distribution of the pattern loop and flow are discussed.

Correlations Between Pattern Parameters

To determine the relationships among different parameters, we evaluated the performance of the Pearson correlation coefficients of the pattern parameters using a correlation matrix (Figure 3) and further visualized specific pairs of correlations (Figure 4).

Figure 3.

Correlation matrix of pattern parameters.

Figure 4.

Correlation among WN and FN, LN, MPL, and HPL.

The correlation coefficients between FN and LN, FN and WN, and WN and LN all reached a considerably high level of approximately .90, which constitutes evidence of a complete correlation, as shown in Figure 4. Moreover, an increase in the WN is likely to lead to a corresponding increase in the LN and FN. In addition, the variation in FN tended to coincide with the variation in LN in the same orientation.

The correlation coefficient between HPL and WN (.72) signifies a robust and salient positive correlation. This relationship is shown in Figure 4, which can be described as and fitted by both a linear trend and a polynomial curve. Thus, an increase in WN generally generates a consistent increase in the pattern layers. The correlation coefficient between HPL and FN, as well as that between HPL and WN, are significantly positive and reach .65. This phenomenon shows the same direction of variation in the other two parameters, thus confirming a universal rule from the perspective of pattern grammar, in which an extension of sentence length represented by the increasing WN leads to an increment of sentence “depth” implied by increasing HPL, FN, and LN.

The results failed to show any salient positive correlations between MPL and other parameters because the correlation coefficient between MPL and WN was .31, and the correlation coefficients between MPL and FN, MPL and LN, and MPL and HPL have low absolute values of .19, .16, and .09, respectively. Therefore, MPL and the group of FN, LN, and HPL can be regarded as two aspects that reflect sentence structural complexity.

Regression Relationships of “Depth” Parameters

In the former discussion, many traces indicate the potential correlation among FN (number of pattern flows), LN (number of pattern loops), and HPL (highest pattern layers), which are three parameters that describe structural complexity from the perspective of sentence “depth.” The pattern flows and loops are regarded as products of the connection points between different pattern layers. Furthermore, the similarity in the FN-WN linearity and LN-WN linearity indicates possible synergetic effects between FN and LN. The correlation analysis of the three “depth” parameters is presented in Figure 5 and Table 7 as follows.

Figure 5.

Correlation between FN and LN, HPL and LN, and HPL and FN.

Table 7.

Regression Analysis of FN and LN, HPL and LN, and HPL and FN.

	Parameter	Coef.	SE	t-Statistics	p > \|t\|	[0.025	0.975]	R ²
$FN = β_{1} * LN + β_{0} + ε$	$β_{1}$	0.977	0.002	420.361	.000	0.972	0.981	.84
$FN = β_{1} * LN + β_{0} + ε$	$β_{0}$	0.739	0.020	36.486	.000	0.699	0.778	.84
$LN = β_{1} * HPL + β_{0} + ε$	$β_{1}$	1.256	0.007	184.985	.000	1.243	1.270	.49
$LN = β_{1} * HPL + β_{0} + ε$	$β_{0}$	−1.297	0.050	−25.937	.000	−1.395	−1.199	.49
$FN = β_{1} * HPL + β_{0} + ε$	$β_{1}$	1.267	0.008	165.974	.000	1.252	1.282	.44
$FN = β_{1} * HPL + β_{0} + ε$	$β_{0}$	−0.808	0.056	−14.365	.000	−0.918	−0.697	.44

Recalling that both FN and LN exhibit robust correlation coefficients that surpass 90%, a valid linear correlation exists between the two parameters, as indicated by their significant linearity. The regression results for FN and LN in Table 7 further demonstrate this correlation effect. With a satisfactory R² of .84, an accurate quantitative relationship has been established, indicating that sentences in Chinese official documents generally contain the same number of pattern flows and pattern loops. Connecting this with linguistic features, a highly meaningful interpretation is that the overlapping structure represented by pattern flows and nominal phrases with complex pre-positive attributes signified by pattern loops should be assigned equal significance to increase the structural complexity of Chinese official documents. Furthermore, in terms of the parallel definitions of pattern flows and loops, this should be viewed as a correlation instead of a dependency.

Regarding the correlation between LN and HPL, each unit increase in HPL will generate 1.256 more pattern loops on average, as shown in Table 7. Such linearity is confined by an R² of .49, which is relatively low and close to the R² of the LN-WN linear relationship. Under these circumstances, the correlation between the LN and HPL shares similar characteristics with that between the WN and HPL, which is a point set that indicates a polynomial convex function.

For each unit of growth of HPL, FN responded with an average increase of 1.267 units. Considering the R² level at .44 and the coefficient, the synergetic effect between LN and FN is preserved even under the influence of HPL; that is, the LN-HPL relation and the FN-HPL relation exhibit a number of common features.

Regression Relationships Between WN and “Depth” Parameters

Considering WN as an interior property of a sentence, this study further pursues a quantitative relationship between WN and other parameters. WN can be treated as the explanatory variable, while the other “depth” parameters (FN, LN, HPL) are regarded as response variables. Therefore, the single-variable regression equations were established separately for WN and FN, WN and LN, and WN and HPL. Table 8 presents the regression results for these three parameter pairs.

Table 8.

Regression Analysis Between WN and FN, WN and LN, WN and HPL.

	Parameter	Coef.	SE	t-Statistics	p > \|t\|	[0.025	0.975]	R ²
$FN = β_{1} * WN + β_{0} + ε$	$β_{1}$	0.296	0.001	376.049	.000	0.294	0.297	.80
$FN = β_{1} * WN + β_{0} + ε$	$β_{0}$	0.990	0.022	45.130	.000	0.947	1.033	.80
$LN = β_{1} * WN + β_{0} + ε$	$β_{1}$	0.276	0.001	367.423	.000	0.274	0.277	.79
$LN = β_{1} * WN + β_{0} + ε$	$β_{0}$	0.900	0.021	43.066	.000	0.859	0.941	.79
$HPL = β_{1} * WN + β_{0} + ε$	$β_{1}$	0.125	0.001	193.843	.000	0.123	0.126	.52
$HPL = β_{1} * WN + β_{0} + ε$	$β_{0}$	3.967	0.018	221.587	.000	3.932	4.002	.52

With a satisfying R² of .80, the regression result indicates that FN increases by .296 for each unit length increase in WN. In other words, for every increment in WN with a length of more than three units, on average, one more pattern flow emerges in the sentence, which contributes to answering Question 2. In addition, considering the mean value of MPL at 2.32, most of the increase in the number of patterns within a sentence can be viewed as a generator for new pattern flows. Thus, it can be concluded that long sentences in Chinese official documents impart complexity through an increasing number of overlapping phrases, which are regarded as representative of pattern flows in Chinese.

As shown in Table 8, a significant quantitative relationship was observed between the LN and WN (R² = .79). In addition, the regression results report a coefficient of .276, which suggests that LN will increase by one for every four-unit length increase in WN, on average. Similar to the performance of the FN, it shows that most of the increases in the number of patterns within a sentence function as a generator for new pattern loops. In other words, with the extension of sentences, growing complexity is achieved by the frequently increasing LN, which is related to the increased adoption of complicated nominal phrases with prepositive attributes.

An increase in WN enhanced the levels of LN and FN almost equally, which is of equal significance in terms of structural complexity. Specifically, Chinese official documents employ more pattern loops and flows, which augment their complexity. This finding implies that processing overlapping structures and complicated nominal phrases is an effective strategy for improving NLP techniques.

The regression parameters reported in Table 8 quantify the correlation between WN and HPL, showing that each increasing unit of WN leads to a .125 greater pattern layer within a sentence. In other words, a sentence in a Chinese official document will be endowed with one more pattern layer for every eight additional words, on average. However, the dependency relation was confined to an approximate linear relation with an R² of .52 rather than total linearity. The increase in pattern layers caused by extended sentences coincided with the growth of FN and LN. Thus, these pattern parameters connect and jointly reflect the structural complexity.

Distribution of “Depth” Parameters Under Fixed WNs

This section focuses on the skewness and kurtosis under fixed or different WN levels to elucidate WN’s influence of WN on the skewness and kurtosis of the pattern parameters.

First, we concentrated on the distribution of certain pattern parameters under three fixed levels of WN (WN = 10, 20, and 30) because these values achieved good representativeness in the corpus.

The distributions of FN and LN at different WN levels are shown in Figure 6. Concerning the distribution of the FN, all approximate PDFs were endowed with significant left skewness or positive skewness coefficients, as presented in Table 9. This finding indicates an obvious preference for a simple sentence structure when sentence length is fixed, which reconfirms the economy principle of language and Zipf’s law.

Figure 6.

Distribution of FN and LN under fixed levels of WN.

Table 9.

Skewness and Kurtosis of FN and LN Under Fixed Levels of WN.

WN	Skewness (FN)	Kurtosis (FN)	Skewness (LN)	Kurtosis (LN)
WN = 10	0.366	0.020	0.125	−0.309
WN = 20	0.305	0.054	0.033	−0.156
WN = 30	0.200	0.064	0.113	0.129

However, comparing the skewness coefficients under fixed WNs and the general skewness coefficient (Sk = 1.253, mentioned in Section “Skewness and Kurtosis of Pattern Parameters,” Table 6) shows that the general skewness coefficient is significantly larger than its counterparts under a fixed WN. A similar preference for simple structures under different levels of the WN also contributed to the general positive skewness coefficient of the FN. Consequently, the general preference for simple structures can be explained by two aspects: (1) the inclination for relatively less complex structures within sentences of a fixed WN and (2) the preference for structures with a certain complexity within sentences of diverse WNs.

Furthermore, as indicated by the positively skewed distribution of FN under different levels of WN, people were more inclined to use flat structures than slim ones. This finding aligns with Hawkins’ (1994) extrapolation based on phrase structural grammar, as well as the inference of H. Liu and Jing (2016) from the perspective of dependency grammar; that is, flat syntactic structures are more easily processed.

In Table 9, all kurtosis coefficients under fixed WNs are close to 0, which reveals insignificant concentration and clustering and shows that the success-breeds-success principle of language does not perform well under the same level of WN. As a result, this within Chinese official documents is largely achieved by the repetition of structures with similar complexity in sentences of diverse lengths, which is a tendency towards a certain complexity under different levels of WN, as discussed above.

The LN distributions under different WNs are presented in Figure 6 and Table 10. Similar to FN, the positive skewness coefficients of LN validate the extrapolation of Hawkins (1994) and the inference of H. Liu and Jing (2016). The difference between the skewness coefficient of LN under fixed WN and the general skewness coefficient of LN indicates that Chinese official documents’ preference for precise and simple structures can be accounted for by (1) the preference for relatively fewer challenging structures among sentences of similar lengths and (2) the inclination for structures with certain complexity among sentences of different lengths.

Table 10.

Skewness and Kurtosis of HPL Under Fixed Levels of WN.

WN	Skewnesscoefficient	Kurtosiscoefficient
WN = 10	1.069	0.091
WN = 20	0.806	1.017
WN = 30	1.069	2.175

As indicated in Figure 7 and Table 10, unlike the skewness coefficients of FN and LN, HPL’s skewness coefficients are higher than those under the general condition when WN was fixed. Moreover, all distributions of HPL under fixed WNs exhibited features of left-skewed unimodal curves with leptokurtosis characteristics, as indicated by significant positive coefficient values. Both the skewness and kurtosis coefficients showed obvious growth with an increase in WN. Specifically, there is a tendency to adopt simple structures in long sentences as well as an inclination to employ structures with constant complexity.

Figure 7.

Distribution of HPL under fixed levels of WN.

As for the skewness of parameters, Figure 8 shows the approximate linear relationship between WN and the skewness of FN with a significant negative correlation coefficient of −.759 and a p-value less than .01. Figure 9 also shows a significant correlation between WN and the skewness of LN, with a coefficient of −.687 and a p-value of less than .01. In Figure 10, the WN and the skewness of HPL were not significantly correlated (p-value > .01). Therefore, as WN increases, the skewness of FN decreases, which means that in Chinese official documents, the longer the sentence, the less preferred it is to adopt flat and simple structures with fewer pattern flows. Furthermore, the decrease in LN indicated that an increase in sentence length weakened the preference for relatively precise structures.

Figure 8.

Skewness and kurtosis of FN under fixed WN.

Figure 9.

Skewness and kurtosis of LN under fixed WN.

Figure 10.

Skewness and kurtosis of HPL under fixed WN.

In terms of the kurtosis of the parameters, the irregular and zigzag broken lines shown in Figure 8 indicate a significant correlation between WN and the kurtosis of FN (p-value > .05). However, Figure 9 shows the positive influence of WN on the effect of structure and collocation clustering because the significantly positive correlation coefficient implies the growth of LN kurtosis during the increase in WN. This occurs because when sentence length increases, collocation clustering in sentences of Chinese official documents is primarily enhanced through pattern loops instead of pattern flows. This is realized through the frequent use of certain complex nominal phrases in relatively long sentences, which aligns with previous findings on the common appearance of proper nouns regarding political and ethnic topics in Chinese official documents (Yue, 2005). A significantly positive correlation also exists between WN and the kurtosis of HPL, as shown in Figure 10, which reflects the contribution of pattern loops to the enhancement of structure clustering and the concentration effect in long sentences.

The correlation between WN and the skewness of MPL was also not significant (p-value > .01), which indicates that the change in the WN did not affect the skewness of the MPL, as shown in Figure 11. It also shows an insignificant correlation between WN and the kurtosis of MPL. It could be inferred that the patterns summarized from real natural language data are often formed by fixed collocations, which are hardly affected by the external factors, that is, sentence length. Nevertheless, according to Figure 11, the skewness of MPL remains consistently greater than zero, which further proves that most Chinese official document sentences would prefer shorter patterns. In addition, the generally positive kurtosis of MPL reconfirms the clustering and concentration of certain structures and collocations among various sentences, as previously discovered.

Figure 11.

Skewness and kurtosis of MPL under fixed WN.

Conclusions

The research began with the establishment of a self-consistent pattern parameter system, which consisted of (1) FN (number of pattern flows), (2) LN number of pattern loops), (3) HPL (highest pattern layers), (4) WN (number of words), and (5) MPL (mean pattern length). Based on this system, we evaluated the structural and distributional characteristics of pattern parameters through their skewness coefficients, kurtosis coefficients, and correlations with the linguistic data of Chinese official documents. The results show that each parameter has good adaptability to the gamma distribution family or chi-square distribution family, and a clear linear correlation exists between the parameters. As indicated in this study, in Chinese official documents, one pattern flow is related to more than a three-word increase in sentence length. On average, one pattern layer was related to an eight-word increase in sentence length, and one pattern loop was related to a four-word increase in sentence length.

By revealing the distribution and correlation of pattern parameters, this study describes and summarizes the composition of Chinese sentences and elucidates the characteristics of official document registers from the perspective of pattern grammar. Therefore, the findings verify the research hypothesis, reflecting how the pattern grammar system can be treated as an effective theoretical framework for describing the structural characteristics of Chinese sentences.

First, the significantly positive skewness of FN, LN, MPL, and HPL reflects the economy principle of language and Zipf’s law (H. Liu, 2022; Zipf, 1949), which also aligns with G. Fang’s (2001) research on the preference of Chinese official documents for accurate and plain expressions. Second, the significantly positive kurtosis of FN, LN, and MPL reflects the concentration and preference of Chinese official documents for a group of collocations and phrases with fixed structures, which corresponds to the inclination to employ fixed collocations and expressions as concluded by Yue (2005). Third, the property of LN confirms the preference of Chinese official document on prepended complicated attributives, like “Shi-construction” and “Wei-construction” as Lv (1982) and Li (2019) mentioned. Fourth, under different fixed levels of WN, the parameters FN, LN, and HPL always maintain significantly positive skewness coefficients, which objectively reflect the preference for flat structures in Chinese official documents. This is consistent with both Hawkins’ (1994) extrapolation based on phrase structural grammar and the inference of H. Liu and Jing (2016) from the perspective of dependency grammar; that is, flat syntactic structures would be easier to process.

In summary, the effectiveness of pattern grammar in describing Chinese syntactic characteristics is confirmed by its agreement with several unique Chinese official document linguistic characteristics, including a preference for a flat sentence structure, fixed collocations, and complex prepositive attributes. Compared to PSG and DG, pattern grammar exhibits authenticity and linearity, making it suitable and reliable for analyzing Chinese sentences and reflecting their complexity.

Above all, this paper presents an effective and innovative theoretical framework that accurately reflects the characteristics of Chinese official documents. Limitations of the study roots in the limit of its corpus size, as it merely focuses on the official documents register. Therefore, based on this framework, further research on other Chinese registers from a pattern grammar perspective will be both feasible and highly significant.

Footnotes

Declaration of Conflicting Interests

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

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: Project supported by the National Social Science Fund of China (No. 17BYY002).

ORCID iDs

Jie Song

Huiying Cai

Data Availability Statement

Data sharing not applicable to this article as no datasets were generated or analyzed during the current study.

References

Brazil

(1995). The grammar of speech. Oxford University Press.

Chametzky

R. A.

(2003). Phrase structure. In Hendrick

(Ed.), Minimalist syntax (pp. 192–225). Blackwell Publishing.

Chomsky

(1957). Syntactic structures. Mouton.

Chomsky

(1995). The minimalist program. MIT Press.

Fang

(2001). Accounts on demands of utility in the use of office paper language. Applied Linguistics, 3, 79–82.

Fang

(2020). A study of high frequency confusing abstract nouns: A new perspective of pattern grammar. Language Teaching and Linguistic Studies, 4, 49–57.

Feng

(2018). Introduction to Chinese register grammar. Beijing Language and Culture University Press.

Hawkins

J. A.

(1994). A performance theory of order and constituency. Cambridge University Press.

Hudson

(1990). English word grammar. Blackwell Publishing.

10.

Hudson

(2007). Language networks: The new word grammar. Oxford University Press.

11.

Hunston

(2019). Patterns, constructions, and applied linguistics. International Journal of Corpus Linguistics, 24(3), 324–353. https://doi.org/10.1075/ijcl.00015.hun

12.

Hunston

Francis

(2000). Pattern grammar: A corpus-driven approach to the lexical grammar of English Philadelphia, America. John Benjamins.

13.

Hunston

(2019). Patterns, constructions, and local grammar: A case study of ‘evaluation’. Applied Linguistics, 40(4), 567–593. https://doi.org/10.1093/applin/amx046

14.

(2019). Corpus-based research on the language characteristics of reporting document-take the official work report for nearly a decade as an example [Unpublished Master’s Thesis, Central China Normal University].

15.

Lian

(1993). Comparative study of English and Chinese. Higher Education Press.

16.

Liu

(2022). Dependency relation and language networks. The Science Press.

17.

Liu

Jing

(2016). A quantitative analysis of English hierarchical structure. Journal of Foreign Languages, 39(6), 2–11.

18.

Liu

(2015). Research on light verb system in modern Chinese [Unpublished Doctoral Dissertation, East China Normal University].

19.

Liu

(2020). N1 of n2 constructions in academic written discourse: A pattern grammar analysis. Journal of English for Academic Purposes, 47, 100893. https://doi.org/10.1016/j.jeap.2020.100893

20.

(2003). On function words in modern Chinese (Revised ed.). Language & Culture Press.

21.

(1982). Essentials of Chinese grammar. Liaoning Education Press.

22.

Perek

Patten

A. L.

(2019). Towards an English constructicon using patterns and frames. International Journal of Corpus Linguistics, 24(3), 354–384. https://doi.org/10.1075/ijcl.00016.per

23.

Price

D. D. S.

(1976). A general theory of bibliometric and other cumulative advantage processes. Journal of the American Society for Information Science, 27(5), 292–306. https://doi.org/10.1002/asi.4630270505

24.

Qiu

Jin

Wang

(2015). A multi-view Chinese treebank based in dependency grammar. Journal of Chinese Information Processing, 29(3), 9–15.

25.

Sinclair

J. M.

(1991). Corpus concordance collocation. Oxford University Press.

26.

Sinclair

J. M.

(1995). Written discourse structure. In Sinclair

J. M.

Hoey

Fox

(Eds.), Techniques of description (pp. 6–31). Routledge.

27.

Tesnière

(1959). Éléments de syntaxe structurale. Klincksieck.

28.

Wang

(2015). Chinese grammar theory. Zhonghua Book Company.

29.

Wang

Feng

Zhang

(2018). Modern Chinese sentence extended pattern grammar model for natural language processing. Journal of Xiamen University (Natural Science), 57(6), 859–866.

30.

Wang

(2008). On the borderline between grammar and lexis: A review of pattern grammar. Contemporary Linguistics, 10(3), 257–266.

31.

Wen

(1984). Some problems in Chinese word order research. Studies of the Chinese Language, 3, 161–165.

32.

Qiao

(2017). Verb reduplication in translated Chinese: A corpus pattern analysis perspective. Journal of Chongqing Jiaotong University (Social Science edition). 17(4), 102–110.

33.

Yao

(2010). A study of de-structures in modern Chinese from the perspective of generative grammar [Unpublished Doctoral Dissertation, Minzu University of China].

34.

Yuan

Huang

(1999). Chinese syntactic structure disambiguation based on semantic knowledge. Journal of Chinese Information Processing, 13(1), 2–9.

35.

Yue

(2005). Document writing tutorial. Higher Education Press.

36.

Zhan

(2021). Distinction between pattern and construction. Foreign Languages and Translation. 28(2), 39–45

37.

Zhang

(2021). A corpus-based study on Chinese modification patterns of nouns across registers. Glottometrics, 51, 12–38.

38.

Zhang

(2013). Modern Chinese personal pronoun and its related syntactic problems [Unpublished Doctoral Dissertation, Central China Normal University].

39.

Zhao

Gao

Liu

(2009). Automatic parsing of coordination structure in contemporary Chinese. In Sun

Chen

(Eds.), Advances of computational linguistics in China (pp. 148–153). Springer.

40.

Zheng

(2013). The syntax of apposition [Unpublished Doctoral Dissertation, Central China Normal University].

41.

Zhou

Huang

(1999). An improved approach for Chinese parsing based on local preference information. Journal of Software, 10(1), 2–7.

42.

Zhu

(1997). Studies on modern Chinese grammar. Commercial Press.

43.

Zipf

G. K.

(1949). Human behavior and the principle of least effort: An introduction to human ecology. Hafner.