For the Underrepresented in Gender Bias Research: Chinese Name Gender Prediction with Heterogeneous Graph Attention Network

Zihao Pan, Kai Peng, Shuai Ling, Haipeng Zhang

Abstract

Achieving gender equality is an important pillar for humankind's sustainable future. Pioneering data-driven gender bias research is based on large-scale public records such as scientific papers, patents, and company registrations. Since gender information is often missing in relevant datasets, studies rely on tools to infer genders from names. However, available open-sourced Chinese gender-guessing tools are not yet suitable for scientific purposes, which may be partially responsible for female Chinese being underrepresented in mainstream gender bias research. Specifically, these tools focus on character-level information while overlooking the fact that the combinations of Chinese characters in multi-character names, as well as the components and pronunciations of characters, convey important messages. As a first effort, we design a Chinese Heterogeneous Graph Attention (CHGAT) model to capture the heterogeneity in component relationships and incorporate the pronunciations of characters. Our model largely surpasses current tools and also outperforms the state-of-the-art algorithm. We open-source a more balanced multi-character dataset from an official source together with our code, hoping to help future research promoting gender equality.

Introduction

In large-scale social science and gender bias research, scholars often need to infer gender from names in datasets where gender isn't explicitly labeled. While tools for English names are mature, Chinese name-gender prediction remains a significant challenge. Existing tools like Ngender rely on simple Naïve Bayes classifiers that treat characters as isolated units. This approach fails to handle out-of-sample characters, ignores the internal structure of characters, and misses the context of multi-character combinations. To address these gaps, we developed CHGAT (Chinese Heterogeneous Graph Attention), a model that analyzes the complex interplay between character components, pronunciations, and multi-character sequences.

Methodology

The CHGAT network utilizes a heterogeneous graph consisting of four node types: characters, semantic components, phonetic components, and pronunciations (pinyin). This structure allows the model to learn the specific roles components play within a character.

The CHGAT model architecture, integrating graph attention with BERT and Transformer modules.

The architecture consists of three primary attention layers:

  • Node-Level Attention: Learns the importance of neighbors within specific sub-graphs (e.g., identifying which radicals are most defining for a character).
  • Structure Attention: Aggregates information from semantic and phonetic component graphs to create a structural representation.
  • Aggregate Attention: Combines the structural representation with pronunciation data (pinyin) to form a final character embedding.

This graph-based approach runs in parallel with a BERT text encoder. The outputs from both are fused and passed through a Transformer module to capture the sequential relationship between characters in a name before a final classifier predicts the gender.

Linguistic Insights: The Power of Components and Pinyin

A key innovation of CHGAT is its ability to distinguish between the semantic and phonetic roles of character components. In Chinese, about 80% of characters are "picto-phonetic." A shared component does not always imply a shared meaning.

Linguistic complexity in Chinese characters: the component '王' serves different roles (semantic vs. phonetic) in different characters.

As shown in the figure above, the component '王' (jade) acts as a semantic component in '珍' (precious) but as a phonetic component in '旺' (prosperous). Previous state-of-the-art models treated these roles identically; CHGAT distinguishes them. Furthermore, the model incorporates pinyin nodes to capture auditory cues, such as the sound "mei" (associated with '美' or beautiful), which is a strong female indicator.

Datasets and Results

To support this research, we introduced a massive new benchmark dataset of 58,393,173 records covering 560,706 unique first names. Sourced from official company registrations, it offers a much more balanced male-to-female ratio (111.59%) than previous datasets like Ngender (197.28%).

In performance testing, CHGAT achieved an accuracy of 93.62% on the new dataset, significantly outperforming the popular Ngender tool (84.76%) and previous neural models like PBERT and FGAT. Even when trained on lower-quality, unbalanced datasets, CHGAT's architectural design proved fundamentally superior at extracting features from Chinese characters.

Conclusion

CHGAT represents a significant step forward in making Chinese names visible in gender bias research. By modeling the internal structure and pronunciation of characters, the model achieves state-of-the-art accuracy. Future work may involve extending these component-aware representations to other NLP tasks, such as sentiment analysis or semantic parsing, and testing the model's generalizability across the wider Chinese diaspora.

Summary updated

Abstract

Achieving gender equality is an important pillar for humankind's sustainable future. Pioneering data-driven gender bias research is based on large-scale public records such as scientific papers, patents, and company registrations, covering female researchers, inventors and entrepreneurs, and so on. Since gender information is often missing in relevant datasets, studies rely on tools to infer genders from names. However, available open-sourced Chinese gender-guessing tools are not yet suitable for scientific purposes, which may be partially responsible for female Chinese being underrepresented in mainstream gender bias research and affect their universality. Specifically, these tools focus on character-level information while overlooking the fact that the combinations of Chinese characters in multi-character names, as well as the components and pronunciations of characters, convey important messages. As a first effort, we design a Chinese Heterogeneous Graph Attention (CHGAT) model to capture the heterogeneity in component relationships and incorporate the pronunciations of characters. Our model largely surpasses current tools and also outperforms the state-of-the-art algorithm. Last but not least, the most popular Chinese name-gender dataset is single-character based with far less female coverage from an unreliable source, naturally hindering relevant studies. We open-source a more balanced multi-character dataset from an official source together with our code, hoping to help future research promoting gender equality.
BibTeX
@inproceedings{pan2023underrepresented,
  author = {Zihao Pan and Kai Peng and Shuai Ling and Haipeng Zhang},
  title = {{For the Underrepresented in Gender Bias Research: Chinese Name Gender Prediction with Heterogeneous Graph Attention Network}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i12.26688},
}
Figure