Introduction
In large-scale social science and bibliometric studies, gender is a critical variable that is often missing from primary datasets, such as patent filings or academic citations. While gender inference tools for Western names are well-established, Chinese name-gender prediction has historically lagged. Traditional tools like Ngender rely on simple Naive Bayes probabilities, which fail to account for "out-of-sample" characters and the complex internal structure of the Chinese script. This gap has led to the systematic underrepresentation of Chinese women in global gender bias research.
The Challenge of Picto-Phonetic Characters
A primary difficulty in Chinese gender prediction is that roughly 80.5% of characters are picto-phonetic. These characters are composed of internal components that signal either meaning (semantic) or sound (phonetic). Crucially, the same component can play different roles depending on the character it inhabits. For example, the component "王" (jade) acts as a semantic radical in "珍" (precious) but as a phonetic component in "旺" (prosperous).
Previous models often treated these components as functionally identical, losing the nuanced "DNA" that signals gendered naming conventions. Modeling this distinction is a core innovation of the current research.
Methodology: The CHGAT Framework
To address these complexities, the authors propose the Chinese Heterogeneous Graph Attention (CHGAT) network. This model treats characters and their components as nodes in a heterogeneous graph, distinguishing between three specific edge types: Character-Semantic, Character-Phonetic, and Character-Pronunciation (Pinyin).
The architecture employs a three-tiered attention mechanism to process this information:
- Node-Level Attention: Determines the importance of neighboring nodes within a specific graph type.
- Structure-Level Attention: Learns whether semantic or phonetic information is more relevant for a given character.
- Aggregate-Level Attention: Combines structural embeddings with pronunciation data.
These graph-derived embeddings are fused with a BERT text encoder to capture the sequential context of multi-character names, followed by a Transformer encoder and a final classifier.
Data and Experimental Setup
The study introduces a massive, high-quality dataset of 58 million records covering over 560,000 unique first names, sourced from official Chinese company registrations and government services. This dataset is significantly more balanced than the legacy Ngender set, with a Male-to-Female ratio of 111.59% compared to Ngender’s highly skewed 197.28%.
Key Results and Insights
CHGAT achieves a state-of-the-art accuracy of 93.62%, outperforming previous benchmarks like FGAT and Pinyin BERT. Key findings include:
- Sequence Sensitivity: Reversing character order in a name flips the predicted gender in 14.77% of cases, justifying the use of sequence-aware models like BERT over simple character-summing approaches.
- Architecture Robustness: Even when trained on lower-quality datasets, CHGAT’s heterogeneous graph structure consistently outperformed simpler models, proving its ability to capture character nuances effectively.
- Importance of Structure: Ablation studies confirmed that the "Structure Attention" layer—which distinguishes between semantic and phonetic roles—is vital for peak performance.
Conclusion
CHGAT provides a powerful tool for more inclusive global research. By modeling the internal linguistic structure of Chinese characters, it offers a high-precision method for gender inference that respects the unique properties of the Chinese language. Future work may extend this heterogeneous representation to other NLP tasks such as sentiment analysis or machine translation.
Summary updated