Navigation – Plan du site

AccueilTransversalitésData papers2024Intercity relationships between 2...


Intercity relationships between 293 Chinese cities quantified based on toponym co-occurrence

Les relations interurbaines entre 293 villes chinoises quantifiées sur la base de la co-occurrence des toponymes
Relaciones interurbanas entre 293 ciudades chinas cuantificadas a partir de la coocurrencia de topónimos
Wang Tongjing, Zhao Yin, Ziyu Bao et Evert Meijers


Cet ensemble de données présente les relations entre 293 villes chinoises, établies à l'aide d'une méthode de co-occurrence des toponymes. Cette méthode évalue la solidité des liens interurbains en mesurant la fréquence à laquelle les noms de deux villes apparaissent conjointement sur une page web. Les données proviennent du corpus d’avril 2019 des archives web Common Crawl, qui contiennent environ 2,5 milliards de pages web. L'objectif principal de cet ensemble de données est de fournir une nouvelle perspective sur les relations inter urbaines, facilitant ainsi les études sur l'analyse des réseaux de villes. L'objectif principal de cet ensemble de données est d'offrir une perspective inédite sur les relations entre villes, facilitant ainsi les recherches sur l'analyse des réseaux urbains. Il stimule non seulement la poursuite des recherches visant à comparer cette relation innovante entre villes avec d'autres réseaux établis, mais sert également de démonstration en présentant une méthodologie simple pouvant être appliquée à d'autres archives au sein de Common Crawl. Cela ouvre ainsi la voie à des analyses longitudinales sur l'évolution des réseaux de villes.

Haut de page

Notes de l’auteur

Credit author statement
Wang Tongjing
: Conceptualization, Formal analysis, Investigation, Methodology, Visualization; Writing – original draft, Evert Meijers: Conceptualization, Methodology; Writing – Review & Editing; Supervision; Project Administration, Ziyu Bao: Visualization, validation, Zhao Yin: Data collection, Software

Texte intégral

1. Introduction

1Cities maintain varying degrees of relationships with a range of other cities (McCann and Acs, 2011; Pumain, 2021). These relationships shape the urban network system, offering opportunities for synergistic urban development (Meijers and Burger, 2017; Glaeser et al., 2016). The introduction of terms such as ‘regional externalities’ (Parr, 2002), ‘borrowed size’ (Meijers and Burger, 2017; Phelps et al., 2001), and ‘urban network externalities’ (Capello, 2000) underlines the external socioeconomic benefits that can be derived from these city relationships.

2Previous research has acknowledged the role of intercity relationships in improving the urban system efficacy (Van den Berghe et al., 2022; Derudder et al., 2022) and in informing sound regional development policy (Meijers and Burger, 2010). However, many studies predominantly rely on mobility-related proxies, such as trains, metrocards, and mobile signals to represent the strength between cities, or focus solely on large cities, due to constraints in data comprehension and accessibility. While there are some exceptions (e.g. Zhang et al., 2016; Pan et al., 2020), gathering data on diverse forms of city relationships, particularly on a large scale, remains a significant challenge.

3Text mining methods offer a new feasible approach to address these limitations by extracting intercity relationships from the text. An effective and straightforward text mining method is collocation analysis (Mello, 2002), or so-called toponym co-occurrence in geography. This method suggests that the strength of the relationship between cities is associated with the co-appearance of city names in text. In other words, the more frequently two cities are mentioned together in the text, the more strongly they are related. This method has been employed at the scale of the USA, China, and Europe (Fize et al., 2016; Meijers and Peris, 2018; Liu et al., 2014), using a variety of text corpora, such as Wikipedia, newspapers, search engines and archives (Overell and Rüger, 2008; Zhong et al., 2017; Peris et al., 2018), which all show that the relationship strength mirrors reality to some extent.

4Compared to machine learning-based text methods, which classify text with millions of features, this word collocation-based method focuses only on one textual feature: the co-appearance frequency of city names. This simplicity allows for straightforward interpretation of results and avoids reliability issues inherent in “black box” machine learning methods (De Graaf and van der Vossen, 2013). Additionally, collocation analysis allows for easy calibration of parameters, making it well-suited for unstructured text, such as webpages and social media texts.

5While the Internet offers a vast, ever-growing source of unstructured text, an ideal source for collocation analysis, processing such a large text source can be difficult in practice. The immense size of digital text corpora, often in the Gigabyte (GB) or even Terabyte (TB) range, makes the use of conventional approaches challenging and necessitates mastery of cloud and distributed computing techniques.

6The objective of this paper is twofold. First, it aims to develop an efficient and user-friendly method that is able to extract collocation analysis results from a large text corpus, thereby detailing the steps to be taken. Second, it aims to offer a ready-to-use dataset presenting how related cities are (broadly defined), thereby offering a valuable resource for researchers interested in urban systems.

7The empirical focus will be on China, where access to large-scale data on city relationships remains limited. Therefore, our study aims to fill this gap by providing a methodological approach and dataset that can serve as valuable tools in understanding the complex pattern of relationships between Chinese cities.

8The remainder of this paper is organized as follows. Section 2 reviews the relevant literature on intercity relationships and text-mining methods, establishing the context for our study. Section 3 describes the methodology employed for corpus extraction and collocation analysis. Section 4 presents the results of our analysis, highlighting key patterns in intercity relationships among Chinese cities. Finally, Section 5 draws the conclusion of this paper.

2. Background

9Collocation analysis was first proposed by Firth (1957), who used the word collocation pattern to identify common language applications. Initially, linguists applied this method focused on grammatical and lexical usage (Halliday, 1966; Firth, 1968; Greenbaum, 1970). Subsequent research extended the application of collocation analysis to include actual content relationships between collocated words, as later research discovered that keywords could also signify the “aboutness” of text (Scott, 1999), and the connection between words may arise from the content of the text itself (Tognini-Bonelli, 2001) or the author's selectional preference (McEnery, 2006).

10Just considering the co-occurrence of words often faces criticism for its apparent simplicity, arguing that collocation may overlook the actual text content of each word co-occurence. While it is a valid point, Baker et al(2008) advocate an interpretative, rather than merely descriptive, approach to collocation analysis. This is because effective communication does not necessarily requires a context more complex than what is essential for interpretation (Brown and Yule, 1982; Tognini-Bonelli, 2001). As such, collocation analysis helps in “limiting the interpretation to what is contextually appropriate or plausible” (Brown and Yule, 1982, p59), and it does, in practice, enable the systematic analysis of large-scale texts (Baker, 2006). However, to better understand why some words collocate more frequently than others (Brezina, 2018), in-depth reading tools like Critical Discourse Analysis are often used to provide a deeper examination of each collocation occurrence (Baker and Vessey, 2018).

11Geographers have also recognized the potential of collocation analysis early on. Tobler and Wineburg (1971) were pioneers in applying the collocation analysis to geography. They counted the co-occurrence frequency of placenames in the cuneiform tablets of the pre-Hittite Cappadocian towns and, inspired by what would become known as Tobler’s law (Tobler, 1970, p236) --“everything is related to everything else, but close things are more related than distant things”, they reconstructed the settlement system based on those co-occurrences.

12As an empirical method, subsequent research has effectively applied collocation analysis in many different cases, revealing that place name co-occurrence can, to some degree, mirror the strength of connections between places (Devriendt et al., 2011; Meijers and Peris, 2019). This approach aligns with Kitchin’s assertion about big data analysis that “instead of solely testing a theory by analysing relevant data, big data analytics seeks to uncover insights born from the data itself” (Kitchin, 2014, p2). Therefore, collocation analysis is viewed as a practical tool capable of quantifying the strength of relationships between cities as represented in the text.

13Early applications of collocation analysis on geography focused on clarifying toponym ambiguities based on placename co-occurrence on Wikipedia (Overell and Rüger, 2008). More recent work has also tended to explore settlement systems, in particular how cities relate to each other through interactions between people, firms, and institutions through digital newspapers, search engines, Wikipedia, and web archives (Devriendt et al., 2011; Liu et al., 2014; Peris et al., 2020; Tongjing et al., 2023a). These studies generally indicate that toponym co-occurrence networks do reflect real-world spatial interactions to a certain extent. An interesting comparison of toponym co-occurrence data with transportation data was conducted by Lin et al. (2019) for cities in Guangdong, a province in China. They analyzed the toponym co-occurrence network generated by Sina, a Chinese news media outlet, and compared it with bus and rail service networks between these cities. Their results show a strong correlation between the overall connectivity of a city as represented by the toponym co-occurrence network and its connectivity as evidenced by the bus and rail service networks. Another interesting way of using toponym co-occurrence data is to incorporate the collocation results into existing models to increase their accuracy. For instance, Wu et al. (2019) incorporated toponym co-occurrence results to refine the field strength model, thereby improving the identification of a metropolitan area’s hinterland.

3. Data collection and processing

14The primary dataset of this study was obtained from the Common Crawl, a web archive that has periodically crawled the Internet since 2008. Starting in 2017, the archive began crawling the Internet on a monthly basis, providing an up-to-date and longitudinal resource. As of 2019, more than 40 TB have been crawled through monthly using the Web ARChive (WARC) format and stored on Amazon S3. We used the entire Common Crawl text archive from April 2019 for processing and conducting experiments (Common Crawl, 2022). This corpus contains approximately 6.98 TB of uncompressed text, comprising 2.5 billion web pages crawled between April 18th and 26th.

15Processing the entire corpus can be time-consuming and costly. Therefore, we initially used a random 0.1% sample (Dataset A) of the corpus to examine the distribution of the share of Chinese tokens in each URL. Next, we downloaded a random subset of the corpus (44,018 URLs, 114 MB WET file, Dataset B) for manual inspection and local processing. Subsequently, we used the entire Common Crawl archive of April 2019 (Dataset C) for processing and conducting experiments using Amazon AWS cloud computing to obtain the filtered corpus. Lastly, we extracted the co-occurrence results from the filtered corpus from a local server. Our processing steps are summarized in Figure 1.

Figure 1. Summary of step-wise toponym co-occurrence extraction

Figure 1. Summary of step-wise toponym co-occurrence extraction

16Clear Chinese word separation was required for further processing. Chinese words often consist of multiple Chinese characters, but a Chinese sentence is composed of consecutive characters without any clear separation. Accurate separation of Chinese characters requires Natural Language Processing techniques. Here we used a popular term frequency-inverse document frequency (TF-IDF) module named JIEBA for word separation (Sun, 2019). This module can separate Chinese sentences into tokens, which can be loosely considered as words. A token is a term in Natural Language Processing that represents a group of characters in the text as a useful semantic unit for processing.

17Since our studies focused on Chinese cities, we opted to filter out non-Chinese webpages to reduce the total amount of data to process. However, Chinese content is often mixed with English or other languages. We defined Chinese webpages as those in which Chinese tokens constitute a significant proportion.

18After using JIEBA to separate Chinese texts into tokens, we examined the distribution of Chinese tokens in the sample webpage data (Dataset A). As shown in Figure 2, we found two peaks. The left peak indicates a significant number of webpages with few Chinese tokens, while the right peak represents webpages with a majority of Chinese content. We included all webpages in the right peak, where the Chinese token ratio was 55% and higher. This approach considerably reduced the workload for subsequent processing while retaining the most relevant webpages.

Figure 2. Chinese Token Ratio Distribution

Figure 2. Chinese Token Ratio Distribution

19Nonetheless, filtering the entire corpus based on each URL’s Chinese token ratio still proved to be an insurmountable task. We decided to use an alternative approach to determine whether a webpage should be included or filtered out. We discovered that the Chinese token ratio is strongly related to the number of continuous Chinese tokens. Figure 3 displays the relationship between the Chinese token ratio with the longest Chinese tokens.

Figure 3. Relationship between the Chinese token ratio and the number of continuing Chinese tokens

Figure 3. Relationship between the Chinese token ratio and the number of continuing Chinese tokens

20In Figure 3, the red line illustrates the relationship between the Chinese token percentage on the webpage and the longest Chinese sequence. The blue line demonstrates that by setting the threshold at 10 continuous Chinese tokens, we can select almost the same number of URLs with Chinese token ratios higher than 55% as possible, while only a few URLs will have ratios lower than 55%. Choosing values lower than 10 would include more URLs with ratios lower than 55% (and fewer URLs with ratios above 55%). We further verified that the URLs filtered by 10 continuing Chinese tokens and 55% ratio were comparable and 96% of them being the same. This filtering process is much more efficient than calculating Chinese token ratios, as it does not require traversing all content of each URL, and results in very similar URLs being selected. We performed this filtering process on a random file, which is a 114 MB WET file (Dataset B). After filtering out non-Chinese languages, we obtained a corpus of 8.3 Megabytes (MB) WET file, which is about 7.28% of the original file.

21Upon examining the documents in the filtered Dataset B, we discovered that pornographic and gambling content made up a significant proportion of the total documents, biasing our data set and diminishing the importance of more relevant texts. As the objective of this paper is also to develop a relevant text corpus in an affordable way, striking a balance between accuracy and efficiency is necessary. To achieve high processing efficiency, we employed only simple processing methods, excluding machine learning-based approaches. Here, we used keyword filtering due to its high efficiency.

22While keyword filtering can be subjective, it simplifies the censoring process and makes it easier to adapt to other circumstances. In general, it strikes a balance between accuracy and efficiency. To remove pornographic and gambling webpages, we used a dictionary with over 10,000 stars on GitHub that contains over 1000 sexual and 68 gambling words (GitHub, 2022). By using this keyword filtering dictionary, the 8.3 MB corpus was filtered down to a 3.8 MB size WET file.

23To complete the entire task, we used a highly scalable Hadoop-based framework for processing the full Common Crawl corpus, utilizing a 1,080 CPU cluster on the Amazon Elastic Map/Reduce infrastructure. To lower costs, we employed AWS spot instances, which were only about 10% of the regular on-demand price. The price of spot instances varies depending on the AWS region and day of the week. During our study, we launched a cluster of 1 master and 29 spot instances of c4.8 xlarge nodes (36 CPUs each). The total operating time was about 58 minutes and the total cost was approximately $30. The resulting output corpus was about 202 GB.

24As processing on a local server was more convenient than on a cloud server, we downloaded the whole filtered corpus (202 GB) from AWS for subsequent processing. Later processing was conducted on a local server.

25The next step was to filter out webpages that did not contain placenames. Chinese placenames can be ambiguous. The co-occurrence of place names may be inaccurate due to toponym ambiguity issues, which means a place name may also refer to other entities. Wacholder et al. (1997) identified multiple types of toponym ambiguities, but in Chinese, the most common toponym ambiguity falls into three categories:

  • Structural: E.g., Beihai, is a city in Guangxi Province, but also means the “north of the sea”.

  • Semantic: E.g., Hezuo, is a city in Gansu Province, but also means “cooperation”.

  • Referent: E.g., Chaoyang, is a city in Liaoning Province, but is also a district name in Beijing.

26To precisely disambiguate toponyms, it is essential to take the context into account, which, however, would require a lot of processing and is time-consuming. Due to the sheer size of the dataset, we decided to exclude these potentially ambiguous cities. Fortunately, most of the toponym ambiguity occurs at the low administration level (e.g. county, village, and district level). Conversely, cities at the prefecture administrative level or higher generally exhibit fewer risks of toponym ambiguity.

27To reduce the risk of toponym ambiguity, we did a two-step process. Initially, we conducted a manual investigation to identify city names that were evidently ambiguous. City names such as “Baiyin” (translating to “silver”), “Dachang” (“big factory”), and “Dongfang” (“East”) were systematically delisted to enhance the precision of the dataset.

28Subsequently, a gravity model was used to conduct a comparative analysis between the absolute frequencies of co-occurrences and the predicted frequencies as calculated by the model. The gravity model can be formalized as follows:

where Iij is the total relationship, K is the constant, Mi and Mj are the sizes of place i and j, respectively, Dij is the physical distance between the two places, 𝛽1 and 𝛽2 reflects the ability of place i and j to attract flows, 𝛽3 reflecting the rate of increase in the friction of distance.

29We then fit the gravity model to the toponym co-occurrence results in order to obtain an estimated strength of intercity relationships. This established a benchmark against which the empirical co-occurrence data could be evaluated. Subsequent analyses focused on outliers—cases where the observed relationships between cities are much higher than the model’s predictions.

30Such outliers were examined to verify whether their high estimation was attributable to naming ambiguities. For example, strong relationships were initially observed between “Beijing” and “Chaoyang,” and “Shanghai” and “Siping”; however, upon further examination, it was found that “Chaoyang” and “Siping”, two small cities, are also district names in Beijing and Shanghai, respectively, thereby clarifying the source of the apparent anomaly.

31After checking the risk of toponym ambiguity, we selected 293 cities.

32The webpages in this corpus that did not contain Chinese characters or the selected Chinese placenames were excluded. The size of the Chinese corpus was then further reduced from 202 GB to 139.5GB of text, which is about 69% of the filtered Chinese corpus. This confirms Hill’s estimate that 70% of documents contain geographical information (Hill, 2009). The filtered Chinese corpus contains approximately 91 million pages from 1,067 top domains and 1,792,759 domain names. In total, the corpus contains around 110 billion tokens, as calculated by the JIEBA tokenizer.

33On average, in the filtered corpus, the top 100 websites ranked by the Chinaz website have about 290 million tokens and 234,691 pages. The total number of tokens in the top 100 websites is about 26.4% of the filtered corpus. To gain a basic understanding of the content in this corpus, we classified the content distribution of these websites. This categorization was according to the primary content areas listed on the Chinaz website for each of these top 100 sites. The percentage distribution of Chinese website types is shown in Figure 4.

Figure 4. The distribution of content on popular Chinese websites

Figure 4. The distribution of content on popular Chinese websites

34Lastly, we tallied the number of webpages where two placenames co-occur and saved the results as a toponym co-occurrence network dataset. The number of webpages where two placenames co-occur serves as a measure of the strength of the relationship between the respective cities.

4. Results

35In this study, the relationship strength between cities was represented by the frequency of toponym co-occurrence. To better understand this data, we approached it from a network perspective, where each city is considered a node and the frequency serves as the link weight between two nodes. Table 1 provides an overview of the toponym co-occurrence network dataset.

36Given the 293 cities in our analysis, there are 42,778 city pairs among which relationships are measured, and we found- toponym co-occurrences for all these pairs, which means it is a full graph.

Table 1. Statistics of the intercity relationships

Type of information


Number of cities


Number of city pairs that co-occur


Mean frequency


Standard deviation


Minimum frequency


First quartile




Third quartile


Maximum frequency


37Table 1 shows that the maximum frequency (7,391,725; Beijing - Shanghai) is more than thirty-two times larger than the third quartile (230,970), while the difference between the third (230,970) and the first quartile (175,844) is relatively small in comparison. The mean frequency (2,272,417) is the average co-occurrence frequency of the 42,778 relationships and is substantially larger than the median frequency (198,880). This substantial difference between the mean and median suggests the presence of a skewed distribution with extreme values. In other words, some city pairs are mentioned far more frequently than others.

38More detailed results can be observed in the distribution of the toponym co-occurrence frequency in Figure 5. This figure presents the distribution of the frequency values on a logarithmic scale (log 10) on the y-axis against the top percentage of relationships on the x-axis. This figure shows that the variation between toponym frequencies is quite significant—90% of the frequency values are between 100,000 and 300,000, while the top 10% extends from 300,000 to 7,391,725. Most of these high values represent relationships between cities with a high administrative level. The most frequent co-occurrence is between Beijing and Shanghai, ranking in the top 0.002% with a frequency of 106.87(7,391,725), and the relationship with the lowest co-occurrence is Naqu-Xianyang, ranking in the top 100% with a frequency of 104.90 (79,985).

Figure 5. The distribution of the toponym co-occurrence frequency

Figure 5. The distribution of the toponym co-occurrence frequency

39Given that the majority of co-occurrence frequency is concentrated within a relatively narrow span, Figure 6 offers a more focused visualization by highlighting the top 10% and bottom 10% of results. This Figure shows that high frequency relationships are primarily situated in the central and southeastern regions of China. Conversely, the low frequency relationships are located in the northern and western areas.

Figure 6. The co-occurrence frequency map of China

Figure 6. The co-occurrence frequency map of China

40To complement Figure 6, Table 2 lists the top 10 intercity relationships with the highest co-occurrence frequency. This highlights that the dominant relationships in China are between Beijing, Shanghai, Chongqing, Tianjin—the four direct-administered municipalities, which are at the highest administrative level, and Guangzhou and Shenzhen, which are among the most prosperous cities following the four direct-administered municipalities.

Table 2. Top 10 intercity relationships with the highest highest co-occurrence frequency


































41Table 3 lists the 10 intercity relationships with the lowest frequency, featuring more dispersed relationships of 12 cities compared to the 6 cities in Table 2. These cities include Naqu—a small city in Tibet—and Macau and Hong Kong, both special administrative regions in China, each appearing in more than one low-frequency relationship. Other cities in this list are Xiangyang, Xining, Liaoyuan, Wuzhong, Qitaihe, Mayannur, Hegang, and Lincang, predominantly small cities located in northern China.

Table 3. 10 intercity relationships with the lowest frequency

























Hong Kong






Hong Kong



42Comparable extreme values can also be observed when analyzing individual cities. We computed the total number of co-occurrence for each city, thus taking into account its relations with all other cities, a measure we refer to as node strength. This measure reflects the extent of each city’s connectivity with other cities. Figure 7 presents the node strength distribution of the general toponym co-occurrence network. The findings suggest that most cities have a total co-occurrence frequency of approximately 60 million times, while only a few cities have been mentioned more than 150 million times.

Figure 7. Node strength distribution of the general intercity relationships

Figure 7. Node strength distribution of the general intercity relationships

43The top 10 cities with the highest overall co-occurrence frequency are presented in Table 4, where Beijing has its dominant presence in the network with the highest co-occurrence frequency of more than 160 million. This is followed by Shanghai, Shenzhen, Guangzhou, and Chongqing with frequencies over 100 million, highlighting their vital interconnectedness in this network. Further down the list are Chengdu, Tianjin, Hangzhou, Wuhan, and Nanjing, whose overall co-occurrence frequency is close to 100 million. The central role of these cities, as highlighted by the data, possibly reflects their economic, cultural, or political significance in the broader scenario under study. A more detailed analysis of how a city’s position in the network associates with its performance can be found in the referenced paper (Tongjing et al., 2023b).

Table 4. The top 10 cities with the highest total co-occurrence frequency


Total co-occurrence frequency





















5. Conclusion

44While it is widely recognized that intercity relationships play a significant role in the urban system, gathering such relational data remains challenging, and current approaches are often skewed toward accessibility or transportation. The collocation analysis method has potential to provide an alternative solution, as it facilitates the retrieval of a new type of large-scale relational data. Nonetheless, urban researchers using the collocation analysis method often encounter difficulties in processing terabyte-scale data, particularly when it necessitates the use of cloud computing. This study showcases an efficient and relatively straightforward approach to extracting intercity relationships from the vast amounts of unstructured text stored in the Common Crawl web archive.

45From an empirical perspective, this study demonstrates how to extract pertinent information via cloud computing, whilst keeping costs to a minimum. The final operation was completed in less than an hour, costing under $100. By deploying a 1080 CPU cluster on the Amazon Elastic Map/Reduce infrastructure, a 202 GB single-language corpus was successfully isolated from a 6.98 TB dataset via a cloud parallel computing method on AWS.

46The toponym co-occurrence method presents several advantages, both practical and theoretical. On the practical front, this method allows for collecting large-scale city relational data that can be difficult to obtain from location data sources, such as mobile signals and metrocards in and out. This makes it a particularly useful method for researchers and policymakers interested in comprehensive, wide-ranging analyses. Theoretically, the method captures a new dimension of city relationships that is distinct from what is revealed through interaction (e.g. transportation) data. While transportation data primarily focuses on the tangible, material movements between cities—such as the flow of goods, services, or people—the toponym co-occurrence method sheds also light on the symbolic or representational associations between cities. These could include cultural, historical, or social linkages that are not necessarily tied to physical movement but are nonetheless significant in understanding the multifaceted relationships between cities. As such, ‘relations’ is a more broad concept than ‘interactions’.

47By extracting relationships from this large text corpus, we ensure a comprehensive and generalized overview of inter-city relationship patterns. However, we acknowledge that our dataset offers merely a temporal snapshot, subject to potential variations when extended to different time frames, languages, and specific genres of text. Also note that the Common Crawl archive is sample data – when taking two consecutive monthly data dumps, about 20% of the urls is similar (CommonCrawl, 2023). These potential variations should not be considered as limitations but opportunities. Such variations underscore the dynamic and evolutionary nature of inter-city relationships in text and pave the way for further investigations in this field.

48While we are optimistic about the potential of the collocation analysis method in general, and more specifically the toponym co-occurrence method in revealing city networks, we acknowledge that the approach is still in its infancy. As such, by presenting our method and dataset, we aim to encourage further engagement on this topic. We have already implicitly suggested several useful future research avenues, including comparisons with interaction data, and an analysis through time. The latter is perfectly possible utilizing the same text source at multiple points in time. Very promising in this regard is the increasing availability of digitalized historic text corpora. Additionally, we call for efforts to classify the types of relationships extracted through collocation analysis, as well as explorations of its wider applicability. Our approach also allows for comparative analyses across cultures or linguistic communities, to see how the perception of relations between cities differs. Finally, while our dataset covers the whole of China, it is obviously also possible to explore regional urban systems using a subset of the cities included.

Dataset description

49Related Publication:
Tongjing, W., Meijers, E., Bao, Z., & Wang, H. (2023b). Intercity networks and urban performance: a geographical text mining approach. International Journal of Urban Sciences, 1-22.

English and Chinese

51Time Period Covered:
2019 April

52Kind of data:
Text tables

53Data Source:
Common Crawl,​the-data/​get-started/​ (accessed 4 June 2023)

54Geographical Coverage:
293 cities in Mainland China

55Geographical Unit:
Cities at the prefecture administration level or above

56Type of article:
Data paper

57Repository location:
Cybergeo Dataverse warehouse:​10.7910/​DVN/​O0M59W

58Direct URL to data:
Wang Tongjing; Yin Zhao; Ziyu Bao; Evert Meijers, 2023, "Replication Data for: Intercity relationships between 293 Chinese cities quantified based on toponym co-occurrence",​10.7910/​DVN/​O0M59W, Cybergeo Harvard Dataverse, V1, UNF:6:cXbLbcdT5vxqR7ETHo53DQ== [fileUNF]

Haut de page


Antaki C., Billig M., Edwards D., Potter J., 2003, "Discourse Analysis Means Doing Analysis: A Critique of Six Analytic Shortcomings", Discourse Analysis Online, Vol.1, No.1, [online]. URL:

Baker P., Gabrielatos C., Khosravinik M., Krzyżanowski M., McEnery T., Wodak R., 2008, "A Useful Methodological Synergy? Combining Critical Discourse Analysis and Corpus Linguistics to Examine Discourses of Refugees and Asylum Seekers in the UK Press", Discourse & Society, Vol.19, No.3, 273-306.

Baker P., Vessey R., 2018, "A Corpus-Driven Comparison of English and French Islamist Extremist Texts", International Journal of Corpus Linguistics, Vol.23, No.3, 255-278.

Baroni B., Bernardini S., 2004, "BootCaT: Bootstrapping Corpora and Terms from the Web", in Proc. 4th International Conference on Language Resources and Evaluation, Lisbon.

Baroni M., Bernardini S., Ferraresi A., Zanchetta E., 2009, "The WaCky Wide Web: A Collection of Very Large Linguistically Processed Web-Crawled Corpora", Language Resources and Evaluation, Vol.43, 209-226.

Brezina V., 2018, "Statistical Choices in Corpus-Based Discourse Analysis", in Taylor C., Marchi A. (eds), Corpus Approaches to Discourse: A Critical Review, London and New York, Routledge, 259-280.

Brookes G., Baker P., 2021, Obesity in the News: Language and Representation in the Press, Cambridge University Press.

Campelo C.E.C., de Souza Baptista C., 2009, "A Model for Geographic Knowledge Extraction on Web Documents", Advances in Conceptual Modeling-Challenging Perspectives, ER 2009 Workshops CoMoL, ETheCoM, FP-UML, MOST-ONISW, QoIS, RIGiM, SeCoGIS, Gramado, Brazil, November 9-12, 2009, Proceedings 28, Springer Berlin Heidelberg, 317-326.

Common Crawl, 2023a, [online]. URL: Accessed June 15, 2023.

Common Crawl, 2023b, [online]. URL:, accessed November 7, 2023.

Derudder B., Feng X., Shen W., Shao R., Taylor P.J., 2022, "Connections between Asian and European World Cities: Measurement, Analysis, and Evaluation", Land, Vol.11, No.9, 1574.

Devriendt L., Boulton A., Brunn S., Derudder B., Witlox F., 2011, "Searching for Cyberspace: The Position of Major Cities in the Information Age", Journal of Urban Technology, Vol.18, No.1, 73-92.

Ediger D., Jiang K., Riedy J., Bader D.A., Corley C., Farber R., Reynolds W.N., 2010, "Massive Social Network Analysis: Mining Twitter for Social Good", 2010 39th International Conference on Parallel Processing, IEEE, 583-593.

Egbert J., Wizner S., Keller D., Biber D., McEnery T., Baker P., 2021, "Identifying and Describing Functional Discourse Units in the BNC Spoken 2014", Text & Talk, Vol.41, No.5-6, 715-737.

Firth J., 1957, "A Synopsis of Linguistic Theory, 1930-1955", Studies in Linguistic Analysis, 10-32.

Firth J., 1968, Selected Papers of J.R. Firth 1952-1959, (F.R. Palmer, Ed.), London, Longmans.

Fize J., Moncla L., Martins B., 2021, "Deep Learning for Toponym Resolution: Geocoding Based on Pairs of Toponyms", ISPRS International Journal of Geo-Information, Vol.10, No.12, 818.

Glaeser E.L., Ponzetto G.A., Zou Y., 2016, "Urban Networks: Connecting Markets, People, and Ideas", Papers in Regional Science, Vol.95, No.1, 17-59.

Greenbaum S., 1970, "Verb-Intensifier Collocations in English", Janua Linguarum, Series Minor 86, The Hague, Mouton.

Ghani R., Jones R., Mladenić D., 2001, "Mining the Web to Create Minority Language Corpora", Proceedings of the International Conference on Information and Knowledge Management, 279-286.

GitHub repository, 2022, "Jieba", [online]. URL: Accessed Sept 14, 2022.

Halliday M.A., 1966, "Lexis as a Linguistic Level", In Memory of JR Firth, 148-162.

Hill L.L., 2009, Georeferencing: The Geographic Associations of Information, MIT Press.

Karami A., Lundy M., Webb F., Dwivedi Y.K., 2020, "Twitter and Research: A Systematic Literature Review Through Text Mining", IEEE Access, Vol.8, 67698-67717.

Lewis S.C., Westlund O., 2015, "Big Data and Journalism: Epistemology, Expertise, Economics, and Ethics", Digital Journalism, Vol.3, No.3, 447-466.

Lin J., Wu Z., Li X., 2019, "Measuring Inter-City Connectivity in an Urban Agglomeration Based on Multi-Source Data", International Journal of Geographical Information Science, Vol.33, No.5, 1062-1081.

Liu Y., Wang F., Kang C., Gao Y., Lu Y., 2014, "Analyzing Relatedness by Toponym Co‐Occurrences on Web Pages", Transactions in GIS, Vol.18, No.1, 89-107.

McCann P., Acs Z.J., 2011, "Globalization: Countries, Cities and Multinationals", Regional Studies, Vol.45, No.1, 17-32.

McEnery T., 2004, Swearing in English: Bad Language, Purity and Power from 1586 to the Present, Vol.1, Routledge.

Meijers E.J., Burger M.J., 2010, "Spatial Structure and Productivity in US Metropolitan Areas", Environment and Planning A, Vol.42, No.6, 1383-1402.

Meijers E.J., Burger M.J., 2017, "Stretching the Concept of ‘Borrowed Size’", Urban Studies, Vol.54, No.1, 269-291.

Meijers E.J., Peris A., 2019, "Using Toponym Co-Occurrences to Measure Relationships Between Places: Review, Application and Evaluation", International Journal of Urban Sciences, Vol.23, No.2, 246-268.

Mello R.A., 2002, "Collocation Analysis: A Method for Conceptualizing and Understanding Narrative Data", Qualitative Research, Vol.2, No.2, 231-243.

Minaee S., Kalchbrenner N., Cambria E., Nikzad N., Chenaghlu M., Gao J., 2021, "Deep Learning-Based Text Classification: A Comprehensive Review", ACM Computing Surveys (CSUR), Vol.54, No.3, 1-40.

Pan F., Bi W., Liu X., Sigler T., 2020, "Exploring Financial Centre Networks Through Inter-Urban Collaboration in High-End Financial Transactions in China", Regional Studies, Vol.54, No.2, 162-172.

Parr J.B., 2002, "Agglomeration Economies: Ambiguities and Confusions", Environment and Planning A, Vol.34, No.4, 717-731.

Peris A., Meijers E.J., Van Ham M., 2018, "The Evolution of the Systems of Cities Literature Since 1995: Schools of Thought and Their Interaction", Networks and Spatial Economics, Vol.18, 533-554.

Peris A., Faber W.J., Meijers E., Van Ham M., January 2020, "One Century of Information Diffusion in the Netherlands Derived from a Massive Digital Archive of Historical Newspapers: The DIGGER Dataset", Cybergeo: European Journal of Geography, [online]. URL:

Phelps N.A., Fallon R.J., Williams C.L., 2001, "Small Firms, Borrowed Size and the Urban‐Rural Shift", Regional Studies, Vol.35, No.7, 613-624.

Pumain D., 2021, "From Networks of Cities to Systems of Cities", in Neal Z., Rozenblat C. (Eds.), Handbook of Cities and Networks, Cheltenham, Edward Elgar Publishing, 16-40.

Scott M., 1999, "WordSmith Tools Help Manual", Version 3.0, Oxford, UK, Mike Scott and Oxford University Press.

Song Y., Shi S., Li J., Zhang H., 2018, "Directional Skip-Gram: Explicitly Distinguishing Left and Right Context for Word Embeddings", Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Vol.2, 175-180.

Sun J., 2012, "Jieba Chinese Word Segmentation Tool", [online]. URL:, accessed November 13, 2022.

Tobler W.R., 1970, "A Computer Movie Simulating Urban Growth in the Detroit Region", Economic Geography, Vol.46, Sup1, 234-240.

Tobler W., Wineburg S., 1971, "A Cappadocian Speculation", Nature, Vol.231, No.5297, 39-41.

Tongjing W., Meijers E.J., Wang H., 2023a, "The multiplex relations between cities: a lexicon-based approach to detect urban systems", Regional Studies, Vol.57, No.8, 1592-1604.

Tongjing W., Meijers E.J., Bao Z., Wang H., 2023b, "Intercity networks and urban performance: a geographical text mining approach", International Journal of Urban Sciences, 1-22.

Tognini-Bonelli E., 2001, Corpus Linguistics at Work, Amsterdam, Benjamins.

Van den Berghe K., Peris A., Meijers E.J., Jacobs W., 2023, "Friends with Benefits: The Emergence of the Amsterdam-Rotterdam-Antwerp (ARA) Polycentric Port Region", Territory, Politics, Governance, Vol.11, No.2, 301-320.

Wacholder N., Ravin Y., Choi M., March 1997, "Disambiguation of Proper Names in Text", Fifth Conference on Applied Natural Language Processing, 202-208.

Wu J., Feng Z., Zhang X., Xu Y., Peng J., 2020, "Delineating Urban Hinterland Boundaries in the Pearl River Delta: An Approach Integrating Toponym Co-Occurrence with Field Strength Model", Cities, No.96, 102457.

Zhang W., Derudder B., Wang J., Shen W., Witlox F., 2016, "Using Location-Based Social Media to Chart the Patterns of People Moving Between Cities: The Case of Weibo-Users in the Yangtze River Delta", Journal of Urban Technology, Vol.23, No.3, 91-111.

Haut de page

Table des illustrations

Titre Figure 1. Summary of step-wise toponym co-occurrence extraction
Fichier image/png, 185k
Titre Figure 2. Chinese Token Ratio Distribution
Fichier image/png, 54k
Titre Figure 3. Relationship between the Chinese token ratio and the number of continuing Chinese tokens
Fichier image/png, 68k
Fichier image/png, 3,2k
Titre Figure 4. The distribution of content on popular Chinese websites
Fichier image/png, 50k
Titre Figure 5. The distribution of the toponym co-occurrence frequency
Fichier image/png, 52k
Titre Figure 6. The co-occurrence frequency map of China
Fichier image/png, 1,0M
Titre Figure 7. Node strength distribution of the general intercity relationships
Fichier image/png, 111k
Haut de page

Pour citer cet article

Référence électronique

Wang Tongjing, Zhao Yin, Ziyu Bao et Evert Meijers, « Intercity relationships between 293 Chinese cities quantified based on toponym co-occurrence », Cybergeo: European Journal of Geography [En ligne], Data papers, document 1057, mis en ligne le 08 janvier 2024, consulté le 26 février 2024. URL : ; DOI :

Haut de page


Wang Tongjing

Department of Human Geography and Spatial Planning, Faculty of Geosciences, Utrecht University, Utrecht, The Netherlands

Corresponding author

Email address:
Twitter handle: TongjingW

Zhao Yin

Department of Intelligent Systems, Faculty of Electrical Engineering, Mathematics, and Computer Science, Delft University of Technology, Delft, The Netherlands

Ziyu Bao

Department of Intelligent Systems, Faculty of Electrical Engineering, Mathematics, and Computer Science, Delft University of Technology, Delft, The Netherlands

Evert Meijers

Department of Human Geography and Spatial Planning, Faculty of Geosciences, Utrecht University, Utrecht, The Netherlands

Articles du même auteur

Haut de page

Droits d’auteur


Le texte seul est utilisable sous licence CC BY 4.0. Les autres éléments (illustrations, fichiers annexes importés) sont « Tous droits réservés », sauf mention contraire.

Haut de page
Rechercher dans OpenEdition Search

Vous allez être redirigé vers OpenEdition Search