(1) Overview
Repository location
Context
This data paper aims to provide additional information for the underlying dataset of the Diary on a Map website,1 a research project which visualizes locations mentioned in Fang Fang’s (方方) Wuhan Diary (Wuhan riji, 武汉日记) on several maps. The project is part of a course on Chinese technical and primary sources in the humanities taught by Dr. Merle Schatz at the University of Göttingen during the summer term in 2025. Fang Fang is a Chinese author and poet. During the outbreak of COVID-19, she started writing a blog on Weibo (微博), a Chinese social media platform, about her daily life and the subsequent lockdown in Wuhan, totaling 60 entries between January 25, 2020, and March 24, 2020 (Xia, 2022). Later, the posts were collected, edited, translated, and published by Michael Berry as a book under the English title Wuhan Diary: Dispatches from a Quarantined City (Fang, 2020). Her blog posts were viewed by millions in China, initially receiving mostly positive feedback but later with increased criticism, especially after the publication of German and English translations (Xia, 2022; Fang, 2020); however, the Chinese Studies scholar Xia (2022) argues that Fang’s insights are generally considered “valuable for the globe’s understanding of our shared pains and worries of this time” (p. 88). With nearly five million followers on Weibo in 2020, Fang Fang can be considered an opinion leader during the pandemic (Tao, 2021).
(2) Method
The following introduces the methodological approach and outlines the different steps for creating the dataset.
Steps
First, the posts were systematically formatted and converted into text files. The formatted files are not part of the dataset due to potential copyright issues, but links to the respective publications are provided. Next, the locations mentioned by Fang Fang in the formatted files were automatically identified through Named Entity Recognition (NER). NER is a Natural Language Processing (NLP) task that aims to identify and classify entities mentioned in texts (Gius, 2024). The primary categories are persons (e.g. public figures, fictional characters), locations (e.g. cities, countries), and organizations (e.g. companies, institutions) (Gius, 2024). Applying NER to Chinese texts requires specifically trained models for the task due to the particularities of the language (Levow, 2006; Gius, 2024). The Microsoft Research Asia (MSRA) NER dataset is often used as a benchmark for evaluating the accuracy of a model (Levow, 2006; Liu et al., 2024). It contains more than 50,000 annotated sentences and more than 80,000 entities (persons, locations, and organizations) from news articles (Jiang & Chen, 2025). It is considered a suitable benchmark when “dealing with formal texts on social media” (Qi et al., 2024, p. 12). The model used for processing Fang Fang’s posts is ERNIE-Gram. It is a pre-trained transformer model that uses an n-gram masking method and has an F1 score of 95.3 on the MSRA NER dataset (Xiao et al., 2021).
Although the model was released in 2021, it still outperforms many newer models (see Table 1). Some models have also achieved marginally better scores, but because these models are not openly available and an F1 score of 95 is generally considered nearly perfect (Gius, 2024), ERNIE-Gram was selected after a manual verification of random sample sentences further confirmed its accuracy. The model is integrated into the workflow through HanLP,2 a Python library for multilingual NLP. Along with entities classified as locations, organizations are also included in the dataset because initial tests showed that in some cases, they should be treated as locations. For instance, although hospitals and universities are identified as organizations, they can usually be resolved to a specific location. Detected locations and organizations were saved to a JSON file together with their daily mention counts and the first sentence in which they appeared for additional context.
Quality control
In the next step, the NER results were manually reviewed. This was necessary for two reasons. First, although ERNIE-Gram generally produces robust results, false positives are inevitable, especially because not all organizations can or should be included on the maps. Second, in order to visualize locations on a map, the coordinates for each place had to be added to the dataset. To simplify this process, all unique locations across the 60 posts were extracted from the raw NER results into a separate file that could then be manually reviewed and edited before the changes were merged back with the raw NER results from each day to produce the final dataset. This approach also has the additional advantage of facilitating contributions to the project. If a mistake is found in the dataset, the data for a location can be corrected in the file containing all unique locations, after which the changes are automatically propagated to the main dataset. Finally, Plotly,3 a Python graphing library, was used to visualize the dataset on several maps. See also Figure 1 for an overview of the files used to produce the dataset.
Figure 1
Data processing flowchart.
(3) Dataset Description
Repository name
Zenodo
Object name
Wuhan Diary Locations: Data and Source Code
Format names and versions
.json, .csv, .html, .geojson, and .py
Creation dates
Between 2025-05-26 and 2026-02-19
Dataset creator
Thore Tyborski
Language
English and Chinese
License
The code and the data are licensed under the MIT license. The .geojson file and the maps are based on data from OpenStreetMap and therefore licensed under the ODbL license.
Publication date
2026-03-11
Details
The repository has three main directories: data, code, and maps. As described above, the Python files in code were used to create the results in the data directory; the maps directory contains the interactive HTML maps. locations_to_map.json, the underlying dataset of these maps, consists of several nested JSON dictionaries. For each post, a dictionary item was created using the respective date as the key. Each date contains a link to the original blog post, the post title, and a dictionary of all locations mentioned that day, with location names used as keys. The different location parameters are presented in Table 2. For convenience, the JSON data has also been converted into a CSV file.
Table 2
Description of location parameters in the dataset.
| DICTIONARY ITEMS | DESCRIPTION | |
|---|---|---|
| KEY | DATA TYPE | |
| mentioned_this_day | Integer | The number of times the location was mentioned on this day. |
| mentioned_total | Integer | The number of times the location was mentioned in total. |
| label | String, English | The NER classification. |
| context | String, Chinese | The first sentence in which the location was mentioned on this day. |
| valid | Nullable boolean | Whether the entry is a valid mappable location. True: shown on the map. False: not a valid location. Null: potentially a location, but not yet resolvable. |
| wuhan | Nullable boolean | Whether the location is in Wuhan. |
| area | Nullable boolean | Whether the location is an area, e.g. a district. |
| lat | Float | The latitude coordinates of the location. |
| lon | Float | The longitude coordinates of the location. |
| description | String, English | A short description or additional information about the location. |
| info_source | String, link | A link to the source of the coordinates. |
(4) Reuse Potential
The reuse potential of the dataset presented here can be considered on three levels: as a source for content-related research, as a workflow applicable to other datasets, and as a resource for evaluating the accuracy of toponym resolution models.
First, the locations serve as a starting point for further research and the identification of points of interest. The parameters in Table 2 allow a fine-grained retracing of the shifting spatial focus over time to reveal how it was affected by certain events. Moreover, the dataset has applications in understanding the effects of the lockdown on the daily life in Wuhan. Considering Fang’s influence and the controversial nature of her posts, the data in combination with other sources could also contribute to the existing literature on how opinion leaders impacted the public discourse on social media (Yi et al., 2022; Yin et al., 2020). Further research could, for example, examine the relationship between the locations mentioned in Fang Fang’s posts and the locations subsequently discussed by users responding to her posts, in order to assess whether her posts shaped the spatial focus of the ensuing discourse.
Second, the workflow introduced here is itself reusable beyond this specific dataset. For instance, Hu et al. (2020) collected 40 million Weibo posts between December 1, 2019, and April 30, 2020. The locations mentioned by other users in such datasets could be identified and mapped using the same workflow. While NER and the visualization of its results are established techniques, existing semi-automated solutions that combine them, such as MARKUS, are predominantly designed for historical texts and do not scale well (Weerdt, 2020). By contrast, the workflow presented here facilitates the creation of visual spatio-temporal representations of locations in large contemporary Chinese datasets and thus has practical applications in Chinese Studies for identifying and mapping these locations. In such cases, the mentioned_total parameter could be used to analyze the most frequently mentioned locations while maintaining a manageable workload.
Third, the dataset may prove valuable for researchers in digital humanities seeking to evaluate the accuracy of automatic toponym resolution models in the future.
(5) Limitations
Finally, researchers should be aware of some limitations when reusing parts of the project. First, Fang Fang sometimes refers to the same location using different terms. For instance, Wuhan may appear as Wuhan (武汉), Han (汉), or Wuhan shi (武汉市). Depending on the research objective, these variants may need to be combined into a single location. Second, researchers should note that running ERNIE-Gram is computationally expensive. For large datasets, processing the sources requires one or more dedicated GPUs.
Notes
[1] https://wuhandiarymap.github.io/, last accessed February 20, 2026.
[2] https://hanlp.hankcs.com/docs/, last accessed February 20, 2026.
[3] https://plotly.com/python/, last accessed February 20, 2026.
Acknowledgements
I would like to thank Dr. Merle Schatz for her feedback on earlier drafts of this paper, as well as her help with the publication process.
Author Contributions
Thore Tyborski: Conceptualization, Data curation, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review & editing.