Word length plays a crucial role in the development of quantitative linguistics, especially in Köhler’s lexical control circuit. There has been a wealth research into word length studies in different languages including Chinese [1] - [8] , yet some boundary conditions are still not specified clearly [9] [10] [11] [12] [13] . A fundamental problem throughout the investigation of word length is the question if there is a universal model with which word length distributions can generally be theoretically described. To this end, many efforts have been made (see [9] for more).

Recently a unified model of length distribution of any unit in language was suggested ( [9] , p. 5) and the authors assumed that “the relative rate of change of the dependent variable (here the frequency) is proportional to the rate of change of the independent variable (here the length)”, which yield the Zipf-Alecseev’s function y = cx^a+bln(x). In the unified model there are merely differences in the parameters, and the parameters themselves are part of a dynamic system displaying self-regulation. The most significance lies in that if we succeed in applying the formula to any level of linguistic entities, we arrive at an enormous simplification.

In this book ( [13] , p. 17), the author stated that the parameter a in Zipf-Alecseev’s function increases with the age of a language, and its values may differ in different languages. Based on the analyses of the values of parameter a in many different languages, Popescu et al. conclude that “one can see that Indo-European languages have in general a smaller parameter a than the languages of other genetic groups. However, Chinese is an exception.” ( [13] , p. 77)

In this study, we will explore whether the text styles or measurement units of word length influence the value of a in Zipf-Alecseev’s function or not. What is more, since the parameters are part of a dynamic system displaying self-regulation, the dependence of the parameter b on parameter a is also tested.

Specifically, the following questions will be explored in this study.

Question 1: Can the word length distributions of Chinese prose and dialogue texts be modeled by Zipf-Alecseev’s function y = cx^a+bln(x)?

Question 2: Do the parameters in fitting Zipf-Alecseev’s function to Chinese word length distributions display any self-regulation (the dependence of the parameter b on parameter a)?

Question 3: Are the parameters in Zipf-Alecseev’s function sensitive to different measurement units of word length (the potential measurement units of Chinese word length are the character and the component)?

Question 4: Are the parameters in Zipf-Alecseev’s function sensitive to different text styles (which are prose and dialogue texts in our case)?

This paper contains four sections. Section 2 describes the materials and methods used; Section 3 presents the results of fitting Zipf-Alecseev’s function to Chinese word length distributions, as well as the comparisons of the values of parameter a between different text styles and different measurement units of word length; Section 4 concludes this study.

In order to measure the word length in spoken Chinese and written Chinese, we built a dialogue text collection (spoken language) and a prose text collection (written language), with 20 texts respectively. The number of words in each text ranges from 726 to 3792. The spoken language texts come from a TV talk show named “QiangQiang San Ren Xing” (in English Three People) on Phoenix TV from 2013.06 to 2013.09, 5 texts each month and 20 texts in total, in the form of daily conversation. This TV program mainly discusses the current social hot issues. The written language texts come from a well-known Chinese prose journal Selective Prose¹, from 2013.06 to 2013.09, 5 texts each month and 20 texts in total.

We need to explain in detail here that, the word “汉语” (means Chinese) consists of two characters “汉” “语”, and five components: “氵” “又” “讠” “五” “口”. Since there are no natural boundaries between words, word segmentation is needed before measuring word length. Word segmentation involves the definition of the word, which is a difficult problem especially in Chinese. But it is not the issue we will discuss here, in the present investigation we segment words with unified standard. Firstly, we use the ICTCLAS, one of the best Chinese word segmentation software, to segment words automatically. Then we did the manual checking and corrected the errors. Table 1 and Table 2 show the number of characters and words tokens in each text.

After word segmentation, we developed a java program to measure word length. To measure the number of components of a word, we used a list consisting of 20902 characters (CJK Unified Ideographs) with numbers of strokes and components of each character.¹

We used Matlab 2012b to do the fitting work, and the goodness of fitting can be seen from the determination coefficients R². As for the statistical comparisons, we used t-test through SPSS 19, and we set the significance level to 0.05 in this study.

Results of fitting Zipf-Alecseev’s function to Chinese word length distributions. In this part we show the results of fitting Zipf-Alecseev’s function to word length distributions of Chinese prose and dialogue texts, including the parameters and the determination coefficients R². What is more, the dependence of the parameter b on parameter a is tested to see if Chinese word length distributions display any self-regulation.

Table 3 presents the results of prose texts, the word length of which is measured in characters.

Using the data from Table 3, the relation between the parameters a and b in Table 3 is visualized in Figure 1. The existence of this link is a sign of self-regulation.

Table 4 also presents the results of prose texts as in Table 3, but the word length is measured in components.

¹Selected Prose Website: http://swsk.qikan.com.

The relationship between a and b in Table 4 is visualized in Figure 2. The existence of this link is a sign of self-regulation.

Table 5 displays the results of dialogue texts, and the word length is measured in components.

The relation between the a and b in Table 5 is visualized in Figure 3. The existence of this link is a sign of self-regulation.

Table 6 also presents the results of prose texts as in Table 5, but the word length is measured in components.

The relation between the a and b in Table 6 is visualized in Figure 4. The existence of this link is a sign of self-regulation.

It can be concluded from the above results that Chinese word length distributions can be modeled by the Zipf-Alecseev’s function, and the dependence of the parameter b on parameter a is testified.

Base on the analyses above, we conclude that:

1) The word length distributions of Chinese prose and dialogue texts can be modeled by Zipf-Alecseev’s function y = cxa + bln(x).

2) The dependence of the parameter b on parameter a is testified, which means that the parameters in fitting Zipf-Alecseev’s function to Chinese word length distributions display some self-regulation.

3) Different measurement units of Chinese word length lead to different values of parameter a in Zipf-Alecseev’s function.

The parameters in Zipf-Alecseev’s function are not sensitive to different text styles (which are prose and dialogue texts in our case), which means that it may be only sensitive to different language types.

This work is supported by the Education Department of Guangdong Province “Innovative Strong School Project” Youth Innovation Talents Project (Humanities and Social Sciences) (Project Number: 2017WQNCX046).

Chen, H. (2018) Comparison of Word Length Distributions in Spoken and Written Chinese. Open Access Library Journal, 5: e4660. https://doi.org/10.4236/oalib.1104660