An Open Dataset of Chinese Name-to-Gender Associations for Gender Prediction in Broad Scientific Research

Abstract

Identifying an individual’s gender based on names is crucial for many scientific studies on gender issues, but it is often complicated by diverse naming conventions globally. Predicting gender from Chinese names poses particular challenges due to unique naming conventions and limited representation in existing datasets. In this study, we introduce a novel dataset comprising 1,051,891 Chinese names in Chinese characters and 96,797 corresponding names in Pinyin from over thirty million Chinese individuals. This dataset includes the frequency of each name’s usage by men and women. We validate our dataset using two additional datasets for predicting gender and find that it offers broader name coverage and higher predictive precision compared to existing methods. Overall, this dataset serves as an essential resource for advancing research in China’s gender-related studies.

Background & Summary

A notable advancement in the last century is the substantial increase in gender diversity across nearly all professional fields^1–5. Gender diversity transcends issues of fairness and equality, serving as a crucial driver of success, innovation, and the welfare of academic research^6,7, industries^8,9, and society as a whole^10,11. Despite these advancements, women continue to be underrepresented and often receive lower compensation in numerous professions³, particularly in high-skilled areas such as academic research^12–14, research and development (R&D)^15,16, and entrepreneurship¹⁴,¹⁷,¹⁸. Therefore, acknowledging and enhancing diversity has extensive impacts across various sectors.

Accurate gender data is vital for studies on gender diversity. Gender detection tools, such as Genderize.io API (https://genderize.io/)¹⁹ and Gender API (https://gender-api.com/), are commonly used when direct data collection proves impractical. For instance, Zhou et al.¹³ employ Genderize.io API to impute the gender of both the first and last authors of 2,432,806 life sciences articles, identifying substantial gender homophily patterns in forward citations. Similarly, Koffi and Marx use Genderize.io API to investigate the gender gap in the commercialization of scientific researchers²⁰. Generally, gender detection tools utilize extensive, often open-access name repositories and enhance results by incorporating additional information about the cultural context, primarily focusing on family names or countries of origin²¹. Additionally, governmental records on name-to-gender associations can be directly used to assign gender by names. Aneja and Reshef¹⁶ infer inventors’ gender using the distributions of first names obtained from the U.S. Social Security Administration (https://www.ssa.gov/oact/babynames/limits.html)²².

However, it has been documented that the accuracy of gender detection tools is quite limited for non-Western names²¹, particularly for Chinese names²³. As shown in Sebo’s work²³, the error rate of commonly used gender detection tools in predicting the gender of individuals with Chinese given names in Pinyin format ranges from 43% to 94%, rendering them ineffective for this population. The low accuracy of these tools for Chinese names can be attributed to three main challenges. First, the datasets used by these tools contain a disproportionately small number of Chinese names. For instance, only 0.57% of the name data in the Genderize.io API is collected from China (https://genderize.io/our-data/)¹⁹. Second, China’s naming conventions and cultural nuances significantly differ from those in Western countries. Third, gender detection tools rely on Pinyin—the official, internationally accepted phonetic system for Chinese words—to predict gender information. This conversion to Latin letters leads to a loss of information, as two distinct Chinese characters can be represented identically in Pinyin, further reducing accuracy. Notably, few studies have investigated which of these challenges is the most limiting for the gender detection of Chinese names. More critically, it remains uncertain whether gender can be reliably predicted from Chinese names as it can from European names.

The lack of gender data hampers our understanding of gender diversity in China’s high-skilled labor markets, including those of scientists, inventors, and entrepreneurs. Despite considerable attention from academia and the media regarding women’s underrepresentation, the absence of quantitative evidence on gender disparities makes it challenging to evaluate the extent of disadvantage faced by women compared to their male counterparts in these sectors.

To address this gap, this article presents a novel dataset consisting of 1,051,891 Chinese names in Chinese characters and 96,797 corresponding names in Pinyin. The dataset includes the frequency of each name used by men and women, enabling researchers to estimate the likelihood of a given name being female. Our dataset is aggregated from 30,282,623 individuals who are shareholders of Chinese firms registered with the State Administration for Industry and Commerce (SAIC) until 2018²⁴. We further utilize two datasets to validate our dataset’s usability as a gender detection tool. One contains 99,729 principal investigators of research projects funded by the National Natural Science Fund of China (https://www.nsfc.gov.cn/)²⁵. The other is a gender-labeled dataset of 9,800 Chinese teenagers (https://www.researchgate.net/publication/269630594)²⁶.

In summary, we demonstrate that our dataset can reliably serve as a gender detection tool for future gender studies across various dimensions. First, it facilitates the assignment of gender to Chinese names in both Chinese characters and Pinyin format. This capability empowers researchers to conduct nuanced investigations into gender disparities using large-scale data where gender information is not directly available, such as data on scientists and inventors. Second, our dataset serves as a benchmark, providing a robust foundation to authenticate and validate supplementary gender-specific datasets related to Chinese names. In addition to its primary relevance for studying gender diversity in China’s high-skilled labor market, the dataset may also support broader inquiries in the global science of science literature. Because it includes both Chinese character and Pinyin representations of names, it can be applied to large-scale demographic inference in bibliometric studies where names are often the only demographic signal available. This builds upon recent work that explores how name-based gender inference tools can both enable and bias research on academic inequality²⁷, and connects to broader investigations into gender stratification in scientific publishing^28–30. Lastly, the dataset’s scalability is notable, as it is prepared for expansion through collaborative efforts to gather additional comprehensive gender-related information.

Methods

In line with the methodology used by Aneja and Reshef¹⁶, we use governmental records to determine gender in our dataset. The data is provided through a collaboration with Beijing Trident Science and Technology Company and was originally obtained from the State Administration for Industry and Commerce (SAIC)²⁴, which mandates all registered firms to formally submit information about their shareholders. Consequently, we obtained the names and gender information of 30,282,623 individual shareholders of firms registered with the SAIC up to 2018. Importantly, this database contains no personal identifiers, thus safeguarding individual privacy.

Initially, we identify names that incorporate the 82 prevalent two-character surnames (details see our source code), with the remaining characters designated as given names. For names not utilizing these surnames, the first character is recognized as the surname, with the subsequent characters constituting the given name. We exclude any names exceeding four characters in length. From this process, we derive 29,674,387 given names from 30,282,623 individuals. The Python ‘pinyin’ package is employed to convert each Chinese character into its Pinyin representation, utilizing a romanization method (available at https://pypi.org/project/pypinyin/)³¹. The Pypinyin package supports polyphone conversion and provides the most common Pinyin results for inputs containing polyphonic characters. These polyphonic characters have minimal impact on the conversion process from Chinese to Pinyin. In our dataset, only 0.07% of individuals(20,741) have names that include polyphonic characters, which could yield different kinds of Pinyin formats.

Data Records

Data structure

Our dataset is composed of 1,051,891 unique first names in Chinese characters and 96,797 in Pinyin format. These were aggregated from a corpus of 30,282,623 individual names by removing duplicates. Table 1 presents the data types of these datasets.

Table 1.

Data type for records in Chinese Gender dataset.

Table	Index	Format	Short description
CnGender	CnName	String	First name in Chinese.
CnGender	Female	Integer	Number of ‘Female’ individuals using the first name.
CnGender	Male	Integer	Number of ‘Male’ individuals using the first name.
CnGender	Ratio	Float	Proportion of ‘Male’ individuals using the first name.
PinyinGender	PinyinName	String	First name in Pinyin (English).
PinyinGender	Female	Integer	Number of ‘Female’ individuals using the first name.
PinyinGender	Male	Integer	Number of ‘Male’ individuals using the first name.
PinyinGender	Ratio	Float	Proportion of ‘Male’ individuals using the first name.

Data description

Our dataset of 30,282,623 names includes 1,051,891 unique first names in Chinese characters and 96,797 in Pinyin format.

Figure 1 illustrates the name-gender distribution within the Chinese Gender dataset, suggesting that gender ratio could serve as a metric for determining the gender associated with a name. Over 60% of individuals have first names (in Chinese characters) predominantly used by males, with over 90% of these individuals being male (Fig. 1a). Conversely, approximately 20% have first names predominantly used by females, with over 90% of these individuals being female. In other words, it is possible to assign gender to over 80% of individuals using a threshold of 0.9 (female ratio ≤0.1 or ≥0.9). Converting names to Pinyin complicates gender prediction because a single Pinyin can represent multiple Chinese characters. Nevertheless, as shown in Fig. 1b, applying the same 0.9 threshold still allows for gender prediction in over 65% of individuals. Figure 1 also indicates that a small proportion of individuals possess gender-neutral first names, with between 40% and 60% being female: about 4.82% in Chinese characters and 7.66% in Pinyin formats.

Fig. 1.

The distribution of names in the Chinese Gender dataset. (a) presents the distribution of names in Chinese characters and (b) presents that in Pinyin format. Names are divided into ten groups according to their female ratio (ranging from 0 to 1), where a higher ratio indicates a name predominantly used by women. The x-axis indicates the female ratio, while the y-axis indicates the population percentage within each group.

We examine the concentration of male and female names within our dataset. As shown in Fig. 2, the names in Pinyin yield less concentrated distribution than that in Chinese. The top 10% of popular names in Chinese account for 70% of the population, whereas the top 10% of popular names in Pinyin account for 63%. This observation underscores that certain common Chinese names account for a significant proportion of the population within our dataset, which aligns with Chinese naming conventions.

Fig. 2.

The concentration of names in Chinese characters (a) and Pinyin format (b) in the Chinese Gender dataset. The horizontal axis represents the top percent of names, and the vertical axis represents the percentage of the total population.

Technical Validation

Test datasets

We utilize two datasets to assess the applicability of our dataset as a tool for gender detection. The first dataset comprises 99,729 grantees of the National Natural Science Foundation of China (NSFC) projects awarded prior to 2015. The NSFC provides public reports detailing grantees, their institutional affiliations, total awarded funds, participating investigators, research focus, and resultant outputs. For our study, we assume that grantees with identical names, within the same research domain and institution, represent the same individual. We compiled a dataset containing 260,809 projects assigned to 117,605 unique scientists. Gender information for each grantee was manually collected from employer websites, personal web pages, Baidu Baike, and public news outlets. Two independent groups of research assistants were employed for this task. In instances of conflicting gender information for the same individual, we conducted a comprehensive verification process, cross-referencing data to reach a final determination. This rigorous procedure ensures the highest accuracy of our dataset. Ultimately, we gathered gender information for 99,729 grantees, representing 84.87% of the original grantees, covering 41,908 unique names in Chinese characters and 19,976 in Pinyin.

The second dataset comprises 9,800 Chinese teenagers aged 16 to 22 years (born in 1992 to 1998). This dataset has been used by its authors to publish several psychology papers, such as Bao et al.³². The dataset was compiled by the authors from public sources such as rosters of universities, middle schools, and primary schools across various regions in China, ensuring no privacy conflicts. We obtained the dataset from the authors’ website (https://www.researchgate.net/publication/269630594)²⁶. The dataset’s authors collected all data from the public resources (rosters of universities, middle schools, and primary schools across different regions in China) on the internet with no privacy conflicts. We identified 6,639 unique names in Chinese characters and 4,223 names in Pinyin format from the 9,800 teenagers.

It is important to note that the ideal test data would be a random sample from the entire Chinese population. However, to the best of our knowledge, such a dataset does not currently exist. The Grantees dataset can be viewed as representing high-skilled labor, while the Teenagers dataset represents a specific cohort. As naming conventions evolve with socioeconomic factors, utilizing these two datasets may introduce some bias in evaluating our dataset.

Performance metrics

Researchers employing a gender prediction tool must consider two critical parameters: frequency and threshold. Frequency refers to the number of occurrences of a given name in the dataset. It represents how many times the tool has encountered that particular name when associating it with a gender. A higher frequency indicates that the name is well-represented in the dataset, potentially leading to more reliable gender predictions because the algorithm has more data to base its gender association on. Threshold is a parameter that can be set to specify the minimum confidence level required for a gender prediction to be considered valid or acceptable. For example, a threshold of 0.8 means that the gender prediction should have at least an 80% certainty based on the available data. If the prediction confidence is below this threshold, the result might be disregarded or questioned. Using a threshold helps filter out uncertain predictions and increase the reliability of the tool’s output, especially when working with names that might have a lower frequency or are more ambiguous in terms of gender association. Together, these metrics help users interpret and control the quality and reliability of the gender predictions, allowing for more informed decision-making when integrating these predictions into applications or analyses.

We employ two approaches to evaluate the performance of using our dataset for gender prediction. First, we set a frequency of one and a threshold of 0.5 to assign gender; that is, if more than half of the individuals with a given name are female, the name is predicted to be female; otherwise, it is predicted to be male. Each individual, regardless of actual gender, can be predicted into three results: ‘female’, ‘male’, or ‘unknown’. A name is classified as ‘unknown’ if it is not found in our reference dataset, and therefore a gender cannot be assigned.We denote these outcomes as $M_m$, $M_f$, $M_u$, $F_f$, $F_m$, and $F_u$, where, for example, $M_m$ represents the proportion of male individuals predicted as male. Note that $M_m$ and $F_f$ are correct classification groups, $M_f$ and $F_m$ are misclassification groups, and $M_u$ and $F_u$ are non-classification groups. Subsequently, we define the five widely used performance metrics^33–35 as follows:

$$\mathit{errorCoded}=\frac{F_m+M_f+M_u+F_u}{M_m+F_m+M_f+F_f+M_u+F_u}$$ 1

$$\mathit{errorCodedWithoutNA}=\frac{F_m+M_f}{M_m+F_m+M_f+F_f}$$ 2

$$\mathit{naCoded}=\frac{M_u+F_u}{M_m+F_m+M_f+F_f+M_u+F_u}$$ 3

$$\mathit{errorGenderBias}=\frac{M_f-F_m}{M_m+F_m+M_f+F_f}$$ 4

$${\mathit{WeightedError}}_w=\frac{F_m+M_f+w\ast \left(M_u+F_u\right)}{M_m+F_m+M_f+F_f+w\ast \left(M_u+F_u\right)}$$ 5

These metrics are straightforward to interpret. $\mathit{errorCoded}$ accounts for both misclassification and non-classification errors, whereas $\mathit{errorCodedWithoutNA}$ and $\mathit{naCoded}$ focus on misclassification and non-classification errors separately. $\mathit{errorGenderBias}$ evaluates the direction of bias in gender prediction; a negative result suggests that the estimated number of males is higher than the actual number, and vice versa. Lastly, ${\mathit{WeightedError}}_w$ captures the prediction errors by using a weighted parameter w to penalize non-classification.

In the second approach, we examine how different methods of utilizing our tool affect performance. We employ two metrics: coverage and precision.

$$\mathit{coverage}=\frac{M_m+F_m+M_f+F_f}{M_m+F_m+M_f+F_f+M_u+F_u}$$ 6

$$\mathit{precision}=\frac{M_m+F_f}{M_m+F_m+M_f+F_f}$$ 7

$\mathit{coverage}$ measures the proportion of the population that could be predicted by our dataset, while $\mathit{precision}$ reflects the proportion of the population accurately predicted by our dataset. Clearly, $\mathit{coverage}$ is calculated as $1-\mathit{naCoded}$ and $\mathit{precision}$ is calculated as $1-\mathit{errorCodedWithoutNA}$. To differentiate prediction performance between males and females, we further analyze precision across different demographic groups.

$$\mathit{precisio}{n}_m=\frac{M_m}{M_m+F_m}$$ 8

$$\mathit{precisio}{n}_f=\frac{F_f}{M_f+F_f}$$ 9

Results

We select Genderize.io API as our primary baseline for comparison, as it is the most widely used resource for imputing gender information^13,21. By submitting the Pinyin format of each name along with the nationality ‘CN’ to the Genderize.io API, we obtain the gender, count, and gender probability for each name. We collect 87,893 results for 12,241 names in the Grantees dataset and 9,278 results for 3,721 names in the Teenagers dataset. The other method we evaluated is based on the name-to-gender inference model introduced by Buskirk et al.³⁶, which was developed to benchmark demographic prediction tools for computational social science. We used their publicly available Python package, NomQuamGender, which provides probabilistic gender inference for names across multiple languages. This method is trained on a large multilingual dataset and aims to reduce geographic and cultural biases inherent in traditional name-based models.

Prediction errors using a threshold of 0.5

As shown in Table 2a, our dataset outperforms Genderize.io API and NomQuamGender in both the coverage of and the accuracy of classifications. In the Grantees dataset, 31% and 21% of individuals (M_u + F_u) could not be assigned a gender by the Genderize.io API and NomQuamGender. These proportions are much larger than those using our methods, whether using Chinese characters or Pinyin. Furthermore, our dataset assigns a higher proportion of correct genders compared to the Genderize.io API and NomQuamGender. Specifically, 71% (68%) of grantees can be correctly assigned to males and 13% (11%) to females respectively using our dataset in Chinese characters (Pinyin), corresponding to 73,323 (71,957) males and 13,426 (11,821) females. By comparison, the Genderize.io API only correctly assigned 42% of grantees to males and 11% to females (58,386 and 12,073 individuals, respectively), while NomQuamGender correctly assigned 56% of grantees to males and 13% tofemales. A similar pattern is observed in the Teenagers dataset, although the difference is less pronounced.

Table 2.

Confusion matrices (a) and Performance metrics (b) of different methods in the test.

Dataset	Method	Confusion matrices/ Performance metrics
(a)		${\mathit{M}}_{\mathit{m}}$	${\mathit{M}}_{\mathit{f}}$	${\mathit{F}}_{\mathit{m}}$	${\mathit{F}}_{\mathit{f}}$	${\mathit{M}}_{\mathit{u}}$	${\mathit{F}}_{\mathit{u}}$
Grantees	Chinese	0.71	0.03	0.07	0.13	0.04	0.01
	Pinyin	0.68	0.07	0.09	0.11	0.04	0.01
	Genderize.io	0.42	0.09	0.07	0.11	0.27	0.04
	NomQuamGender	0.56	0.14	0.07	0.13	0.07	0.13
Teenagers	Chinese	0.45	0.02	0.17	0.30	0.03	0.03
	Pinyin	0.46	0.03	0.31	0.19	0.00	0.00
	Genderize.io	0.38	0.09	0.16	0.31	0.03	0.02
	NomQuamGender	0.34	0.11	0.07	0.30	0.05	0.03
(b)		$\mathit{error\; Coded}$	$\mathit{errorCoded\; WithoutNA}$	$\mathit{naCoded}$	$\mathit{errorGender\; Bias}$	$\mathit{Weighted}{\mathit{Error}}_{\mathit{w}}$
	Chinese	0.13	0.11	0.02	−0.03	0.11
Grantees	Pinyin	0.16	0.16	0.00	−0.02	0.16
	Genderize.io	0.29	0.20	0.12	0.03	0.22
	NomQuamGender	0.31	0.23	0.10	0.08	0.25
	Chinese	0.24	0.20	0.06	−0.15	0.21
Teenagers	Pinyin	0.35	0.34	0.01	−0.27	0.34
	Genderize.io	0.31	0.27	0.05	−0.09	0.28
	NomQuamGender	0.36	0.30	0.08	−0.05	0.31

We follow the prior work^22,23,34 and set WeightedError w’s weighted parameter w as 0.2. In Table 2(a), values represent the proportion of the total test population in each category. The sum of all six cells in a given row should equal 1. Please note that the sum may not be exactly 1 due to rounding each value to two decimal places.

Table 2b further illustrates that all error metrics for using our dataset to predict gender are smaller than those for the Genderize.io API and NomQuamGender in the Grantees dataset. For instance, the errorCoded metric is 0.13 for Chinese characters and 0.16 for Pinyin with our dataset, both significantly lower than those with the Genderize.io API and NomQuamGender. Our dataset, when using Chinese characters, also outperforms the Genderize.io API and NomQuamGender in most error metrics within the Teenagers dataset. However, the errors associated with using the Pinyin method to predict gender in the Teenagers dataset are higher than those with the Genderize.io API and NomQuamGender. This discrepancy may be due to the Teenagers dataset being drawn from a specific cohort of the Chinese population, whose gender-name distribution differs from that of our dataset.

Coverage and precision using different frequencies and thresholds

As mentioned above, researchers utilizing a gender prediction tool usually take into account two essential parameters: frequency and threshold. Analyzing coverage and precision across different frequencies and thresholds is crucial for gaining a comprehensive understanding of the dataset’s effectiveness.

We first assess the coverage of methods for gender prediction and find that the Pinyin method achieves the highest name coverage as the frequency threshold increases. From a frequency range of 0 to 200, both the Chinese and Pinyin methods surpass the Genderize.io API in terms of name coverage for predictions in the Grantees dataset (Fig. 3a) and Teenagers dataset (Fig. 3b). The Pinyin method demonstrates superior coverage compared to the Chinese method because a single Pinyin corresponds to multiple Chinese character forms. Among the four methods, Genderize.io and NomQuamGender exhibit more pronounced downward trends, largely due to their smaller dataset for Chinese names, whereas the Chinese and Pinyin methods experience a more gradual decline. We place particular emphasis on the frequency of 10, as researchers typically set a general threshold (name frequency ≥ 10) to mitigate the influence of extremely rare names. When the frequency reaches 10, the coverage of Genderize.io and NomQuamGender drop significantly to approximately 50% for the Grantees dataset and 70% for the Teenagers dataset. In contrast, the Pinyin (Chinese) method maintains a high coverage, exceeding 91% (90,290 grantees) and 82% (7,990 teenagers), respectively, for the Grantees and Teenagers datasets. We further investigate how name coverage behaves when the threshold changes. The Fig. 3(c,d) shows that our Chinese Gender methods have better name coverage than Genderize.io and NomQuamGender. As the threshold increases, the number of individual that can be predicted shrinks. At the threshold of 0.8, for grantees, the Chinese methods lead with over 73% coverage, significantly outperforming Genderize.io, which reaches 31%. For teenagers, Genderize and NomQuamGender still give name coverage less than the other two methods. After the threshold is greater than 0.70 and 0.73, the number of names covered by Chinese exceeds that of Pinyin. Genderize.io and NomQuamGender continue to exhibit the lowest coverage until around a threshold of 0.97. We also examine the coverage of female and male names separately for the Grantees and Teenagers datasets (Fig. 4). For both datasets, female name coverage is generally slightly higher than that of male names across the four methods as name frequency increases, with the exception of the Teenagers dataset (Fig. 4a,b). However, when viewed through the lens of classification threshold, female name coverage is lower than male name coverage for the Chinese and Chinese Pinyin methods (Fig. 4c,d), although the overall performance of these two methods remains substantially better than that of Genderize.io and NomQuamGender. In Fig. 4c, at a threshold of 0.8, the coverage of male names in the Grantees dataset using the Chinese Pinyin method is 73%, whereas the coverage of female names is only 48%. A similar pattern is observed in the Teenagers dataset (Fig. 4d), where male name coverage is 68% and female name coverage is 37%.

Fig. 3.

Name coverage in the Grantees (a) and Teenagers (b) over various frequency settings. The x-axis represents the frequency of names, the y-axis represents the proportion of names under a limit of name frequency. The red dotted line is located on frequency = 10. Name coverage in the Grantees (c) and Teenagers (d) over various threshold settings. We show the coverage of the Grantees (c) and Teenagers (d) datasets by setting various thresholds, with a frequency of 10. The x-axis represents the gender threshold (female = 1, male = 0), and the y-axis denotes the proportion of males (=<threshold) and females (>= threshold). Adjusting the threshold determines the confidence level for categorizing a name as male or female, affecting the number of classified individuals.

Fig. 4.

Gender-disaggregated name coverage in the Grantees and Teenagers datasets across varying frequency and threshold settings. (a,b) Name coverage within gender in the Grantees (a) and Teenagers (b) datasets over various frequency settings. The x-axis represents the frequency of names, while the y-axis denotes the proportion of names under a given frequency limit, disaggregated by gender. A red dotted line is located at frequency = 10. (c,d) Name coverage within gender in the Grantees (c) and Teenagers (d) datasets over various threshold settings. These panels show the coverage of names for each gender by setting various gender probability thresholds, with a fixed name frequency of 10. The x-axis represents the gender threshold (female = 1, male = 0), and the y-axis denotes the proportion of males (probability <= threshold) and females (probability >= threshold). Adjusting the threshold determines the confidence level for categorizing a name as male or female, affecting the number of classified individuals for each gender.

We subsequently examine the precision of gender prediction as the threshold varies, as illustrated in Fig. 5. Our findings reveal that the dataset supports more precise gender prediction tools, namely the Chinese and Pinyin methods, compared to Genderize.io and NomQuamGender. The Chinese method exhibits the highest precision, followed by Pinyin, while Genderize.io shows the lowest precision. In the male group of the Grantees dataset (Fig. 5a), the Chinese method outperforms Genderize.io, with all methods achieving approximately 90% precision at initial thresholds and then showing steady improvement. For female prediction (Fig. 5b), Genderize.io’s precision ranges only from about 57% to 73% in the Grantees dataset, NomQuamGender ranges from 8% to 32%, whereas the Chinese method attains accuracy levels of approximately 79% to 94%. Among teenagers (Fig. 5c), the Chinese method again achieves the highest precision within the male group. Notably, Pinyin underperforms relative to Genderize.io when the threshold exceeds 0.8 (Fig. 5c), potentially due to Genderize.io’s low coverage of around 36% (Fig. 4d). For females (Fig. 5d), name prediction yields better results than for males, with precisions of 94%, 85%, and 77% for the Chinese, Pinyin, and Genderize.io methods at the 0.5 threshold, respectively. Our evaluation shows that the prediction accuracy of the test dataset remains comparably high when using Chinese Pinyin method, indicating that the Chinese Pinyin dataset is robust and suitable for name-gender association tasks where only Romanized names are available. Overall, these evaluations further affirm the robustness and utility of the Chinese Gender dataset.

Fig. 5.

The relationship between precision and threshold. Here we restrict the frequency of names to 10 or more and perform the analysis separately for male (a,c) and female (b,d) groups. The upper and lower sides denote the results of the Grantees (a,b) and Teenagers (c,d). The threshold range is 0.5–1, with a 0.01 increment.

Performance with Chinese names outside mainland China

To evaluate generalizability to Chinese names outside mainland China, we used the name-ethnicity-classifier³⁷ to identify names of likely Chinese origin in the U.S. Social Security Administration (USSA) dataset. Among 24,453 individuals’ first names, our Chinese Pinyin method showed reduced coverage and lower precision on this subset (Fig. 6), indicating potential differences in naming conventions among diaspora populations. At the frequency threshold of 10, the Chinese Pinyin method can predict 44.3% names of USSA datasets and the coverage of predictable names is around 28% when the threshold meets 0.8. The predicting results yield 58% and 75% precision for males and females.

Fig. 6.

Coverage and precision of the Chinese Pinyin method on Chinese overseas names (the USSA dataset). (a) Cumulative share of names in the USSA dataset as a function of name frequency (x-axis). The y-axis shows the cumulative proportion of names captured below each frequency, with the red dotted line marking a frequency of 10. (b) Share of names classified in the USSA dataset across different gender probability thresholds (x-axis), with the frequency cutoff fixed at 10. This illustrates how coverage varies with the confidence level required for classification. (c) Precision for male classifications and (d) precision for female classifications across gender probability thresholds (x-axis). The y-axis indicates the proportion of correctly classified male or female names among all names assigned to that gender. Adjusting the threshold affects both the confidence and the number of names that can be classified.

The reduced performance may stem from multiple factors, including cultural differences in naming conventions and the inclusion of non-Chinese names that resemble Pinyin spellings. These findings underscore the importance of region-specific validation and we therefore suggest that future work to consider the development of targeted gender inference tools for Chinese diaspora populations.

Usage Notes

In this study, we present an innovative dataset comprising 1,051,891 Chinese names in Chinese characters and 96,797 corresponding names in Pinyin. The dataset provides information on the frequency of each name’s usage by men and women, allowing researchers to estimate the likelihood of a name being associated with females. Our dataset demonstrates strong performance in gender prediction using both Chinese characters and Pinyin. At a commonly used frequency of 10 and a threshold of 0.8, both Chinese and Pinyin methods achieve substantially higher name coverage (Fig. 3) and gender predicting precision (Fig. 5) compared to Genderize.io and NomQuamGender. We illustrate the feasibility of predicting an individual’s gender based on their Chinese name with appropriate tools.

This dataset represents an initial step in supporting the community’s exploration of gender-related issues. Several promising directions exist for future work to build on this effort. Firstly, although we have validated the performance of the Chinese Gender dataset using the Grantees and Teenagers datasets, further validation using a random sample from the overall Chinese population is needed. Secondly, despite the small proportion of common polyphonic names in the dataset, enhancing conversion accuracy by considering the gender probability of character combinations in such names could improve prediction accuracy.

Limitations

There are some limitations to be considered and noticed when using Chinese Gender dataset. The first is our method is limited to binary gender inference and does not account for non-binary or ethnically diverse identities. This reflects a broader constraint of current name-based demographic tools, as noted in recent work by Lockhart et al.²⁷.

While our dataset substantially improves coverage compared to existing tools such as Genderize.io, it is important to recognize that it is constructed from shareholder records of Chinese companies—a population disproportionately composed of older, male individuals from middle-to-upper socioeconomic backgrounds. As shown in Fig. 4c,d, the coverage of female names is consistently lower than that of male names across methods, particularly at high thresholds. This pattern likely reflects the male-skewed composition of the source data. As a result, certain name patterns may be overrepresented, particularly those associated with these demographic groups. At present, there is no large-scale, demographically representative benchmark dataset of Chinese names with verified gender labels. This limits our ability to evaluate model performance across the full spectrum of the Chinese population.

We encourage future work to address these gaps by developing more inclusive datasets and exploring strategies that can support non-binary, self-identified, and ethnically diverse identities for Chinese names. Enhancing demographic coverage and fairness in name-based inference will be critical for equitable applications in related research.

Author contributions

D.S. conceived the idea, collected the data and designed the experiments; S.T. performed data processing, data analysis and validation with the help of D.S.; S.T. and D.S. wrote and edited the manuscript.The authors contributed equally to this work and are listed in alphabetical order.

Data availability

The dataset, Chinese Gender, is freely available at Harvard Dataverse³⁸: 10.7910/DVN/UAARYY.

Code availability

The source code for data curation, data linkage, and metrics calculation is available at https://github.com/tongt1213/Chinese-Gender-dataset.

Competing interests

The authors declare no competing interests.