CORGI-PM: A Chinese Corpus For Gender Bias Probing and Mitigation

Ge Zhang, Yizhi Li, Yaoyao Wu, Linyuan Zhang, Chenghua Lin, Jiayi Geng, Shi Wang, Jie FuarXiv 2023 · DOI · arXiv

Introduction

Gender bias in Natural Language Processing (NLP) has predominantly been studied through the lens of the English language. Existing Chinese datasets often rely on narrow word-level associations or automated "gender-swapping" techniques, such as replacing male pronouns with female ones. However, these methods frequently fail in Chinese due to deep-seated cultural nuances and social roles; a simple swap can result in nonsensical sentences. To address this, researchers have introduced CORGI-PM, a large-scale, human-annotated Chinese corpus designed for sentence-level gender bias detection, classification, and mitigation.

Methodology

The creation of CORGI-PM involved a multi-stage pipeline designed to capture contextual bias beyond simple keyword matching.

The two-stage "Recall and Filter" pipeline used to construct the CORGI-PM corpus.

The process began with Candidate Retrieval, using a "Recall and Filter" approach. Gender-related vocabulary was identified via word embeddings, and Large Language Models (LLMs) like ERNIE and CBert were employed to assign gender bias scores to sentences from massive raw corpora (SlguSet and CCL).

Following retrieval, six native Chinese speakers performed Human Annotation, labeling sentences and categorizing bias into three types:

  • AC (Activity/Career): Stereotypes regarding occupations or hobbies.
  • DI (Descriptions/Inductions): Stereotypical physical or personality traits.
  • ANB (Attitudes, Norms, and Beliefs): Societal expectations or sexist beliefs.

Finally, for every biased sentence, annotators created a Mitigation Paraphrase—a neutral version that preserves the original intent while removing biased language.

The CORGI-PM Dataset

The resulting CORGI-PM dataset consists of 32.9k sentences. This includes 5.2k biased sentences paired with their human-written neutral counterparts, alongside 22.5k non-biased samples. The dataset is partitioned into standard Train, Validation, and Test splits to facilitate reproducible benchmarking in the research community.

Analysis of Gender Bias

A key finding of the study is the failure of simple gender-swapping. For instance, "cleaning woman" (清洁阿姨) cannot always be swapped to "cleaning uncle" (清洁叔叔) without losing cultural coherence; a neutral term like "cleaner" (清洁工) is more appropriate.

Word clouds illustrating gendered associations for adjectives and careers across different Chinese language models.

Probing popular Chinese language models (LMs) like BERT-wwm and XLNet revealed significant inherent biases. As shown in the word clouds, these models consistently associate specific adjectives and career paths with particular genders, reflecting the biases present in their training data. Interestingly, biased sentences and their corrected versions tend to be more linguistically diverse and longer than neutral sentences, suggesting that expressing or correcting stereotypes requires more complex descriptive language.

Experimental Results

The authors established baselines for two primary tasks:

  1. Detection & Classification: Models such as BERT and RoBERTa achieved F1-scores between 0.55 and 0.82. While detecting the presence of bias was relatively straightforward, classifying the specific type of bias (AC, DI, or ANB) proved more challenging.
  2. Mitigation (Correction): GPT-3 variants were tested on rewriting biased sentences. Fine-tuned models (e.g., Babbage) outperformed zero-shot models on automated metrics like ROUGE-L. However, human evaluation noted that while the zero-shot Davinci model was more conservative in its changes, it often maintained better coherence, even if it missed some subtle biases caught by smaller, fine-tuned models.

Conclusion and Future Work

CORGI-PM provides a robust foundation for addressing gender bias in Chinese NLP. While the current work focuses on binary gender and reflects the perspectives of highly educated annotators, it sets a benchmark for future expansions. Future research should aim to include non-binary identities and explore the performance of even larger-scale models as computational resources allow.

Summary updated

Abstract

As natural language processing (NLP) for gender bias becomes a significant interdisciplinary topic, the prevalent data-driven techniques such as large-scale language models suffer from data inadequacy and biased corpus, especially for languages with insufficient resources such as Chinese. To this end, we propose a Chinese cOrpus foR Gender bIas Probing and Mitigation CORGI-PM, which contains 32.9k sentences with high-quality labels derived by following an annotation scheme specifically developed for gender bias in the Chinese context. Moreover, we address three challenges for automatic textual gender bias mitigation, which requires the models to detect, classify, and mitigate textual gender bias. We also conduct experiments with state-of-the-art language models to provide baselines. To our best knowledge, CORGI-PM is the first sentence-level Chinese corpus for gender bias probing and mitigation.
BibTeX
@misc{zhang2023corgi,
  author = {Ge Zhang and Yizhi Li and Yaoyao Wu and Linyuan Zhang and Chenghua Lin and Jiayi Geng and Shi Wang and Jie Fu},
  title = {{CORGI-PM: A Chinese Corpus For Gender Bias Probing and Mitigation}},
  year = {2023},
  eprint = {2301.00395},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2301.00395},
}
Figure