Search results
CORGI-PM 🐶 is a Chinese cOrpus foR Gender bIas Probing and Mitigation, which contains 32.9k sentences with high-quality labels derived by following an annotation scheme specifically developed for gender bias in the Chinese context.
- 2.2 Annotation Scheme
- 3 Gender Bias Mitigation Challenges
- 4 Conclusion
- Limitations
- Ethics Statement
- A.2 Discussion
- B.1 Word Cloud Analysis
- D Case Study
The annotation scheme is designed for gender bias probing and mitigation. For gender bias probing, the annotators are required to provide the follow-ing information given a sentence: whether gender bias exists; if so, how the bias is established. For gender bias mitigation, the corrected non-biased version of the biased sentences is also required. ...
To provide a clear definition for automatic textual gender bias probing and mitigation tasks, we pro-pose corresponding challenges and standardize the evaluation protocols. We address two tasks, detec-tion, and classification, for gender bias probing and formalize the gender mitigation challenge as a text mitigation task.
We propose CORGI-PM, the first Chinese human-annotated corpus for both gender bias probing and mitigation. We also address definitions and evalua-tion metrics for three challenges based on CORGI-PM and test the performances of state-of-the-art language models. Our proposed challenges can serve as benchmarks for measuring the ability of language mod...
There are several major limitations in this research work. Due to the high requirement of annotators for annotating gender-biased sentences and correct-ing such sentences, we only choose annotators with higher education, which may lead to potential cog-nitive bias. In addition, we only conduct limited implementations and experiments of testing wide...
We carefully consider the ethical implications dur-ing the collection process. The collection of our corpus CORGI-PM sentences only relies on public available corpora for research purposes. We have acknowledged the potential usage of our dataset as well as related privacy issues to the annotators and received confirmations before the annotation was...
There exists observing gender bias in the open-source Chinese language models, especially in Ernie and Chinese Word Vectors according to Fig. 3. We hypothesize that the observation is highly related to the corpus used. Cui et al. claim that their used corpus is a combination of Chine-seWiki, and some other universal Chinese datasets, including ency...
We provide word cloud analysis of Ernie and Chinese-Electra in the section about adjectives and career words. More available word cloud analy-sis will be available in our public repository. The words are ranked according to the absolute value of their gender bias score calculated along the method used by Bolukbasi et al.; Jiao and Luo. There is a n...
As shown in Fig. 6, gender-swapped methods suffer from mitigating gender bias expressed by gender-specific descriptions and inductions, and expressed gender-stereotyped attitudes, norms and beliefs. As a result, gender-swapped methods may generate nonsensical sentences under certain circumstances. We also use the basic mitigation annotation pat-ter...
Jan 1, 2023 · 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.
Articles 1–20. University of Manchester, PhD student - Cited by 244 - Natural Language Processing - Multimodal Learning - Music Modeling.
Jan 1, 2023 · 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.
Jan 1, 2023 · CORGI-PM: A Chinese Corpus For Gender Bias Probing and Mitigation. January 2023. DOI: 10.48550/arXiv.2301.00395. Authors: Ge Zhang. University of Michigan. Yizhi Li. Yaoyao Wu. Linyuan Zhang....
Jan 1, 2023 · This work proposes a Chinese corpus, 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, and addresses three challenges for automatic textual gender bias mitigation.
