Mitigating Language-Dependent Ethnic Bias in BERT

BERT and other large-scale language models (LMs) contain gender and racial bias. They also exhibit other dimensions of social bias, most of which have not been studied in depth, and some of which vary depending on the language.

In this paper, we study ethnic bias and how it varies across languages by analyzing and mitigating ethnic bias in monolingual BERT for English, German, Spanish, Korean, Turkish, and Chinese.

Then we propose two methods for mitigation; first using a multilingual model, and second using contextual word alignment of two monolingual models. We compare our proposed methods with monolingual BERT and show that these methods effectively alleviate the ethnic bias.

Measuring Ethnic Bias

We define ethnic bias in BERT as the degree of variance of the probability of a country name given an attribute in a sentence without any relevant clues.

We formally define normalized probability used in our ethnic bias metric. The metric is based on the change-of-probability of the target words given the presence or absence of an attribute word. If a model predicts more uniform normalized probabilities to all target groups, the model's ethnic bias is lower.

We propose the Categorical Bias (CB) score, defined as the variance of log normalized probabilities. If the model predicts uniform normalized probabilities to all target groups, then the CB score would be 0. On the contrary, for a model with a high ethnic bias, the CB score would also be very high.

The bias is easily seen in the chart above. Depending on the language in which the monolingual BERT is trained, different attributes are predicted for the same question.

Mitigation Result

We propose and apply two solutions: multilingual BERT(M-BERT) and contextual word alignment.

Original M-BERT helps to greatly reduce ethnic bias for English, German, and Spanish. At first, ethnic bias is not mitigated with the M-BERT for Chinese. But in the end, M-BERT with fine-tuning performs the best on the high-resource languages, including Chinese.

We choose to align the monolingual LMs to English BERT, and this mitigation is effective for Korean and Turkish for which the M-BERT mitigation is not very effective. We find the method is useful to mitigate ethnic bias for the languages in which M-BERT performance degradation occurs due to a relative lack of resources.

Comparison Map

Conclusion & Future Work

In this paper, we study language-dependent ethnic biases in BERT. To first quantify ethnic bias, we introduced the category bias (CB) score.

We show the language-dependent nature of ethnic bias, and then we proposed two mitigation strategies: multilingual model and contextual word alignment with English, which has the lowest CB score. For resource-rich languages, the multilingual model alone can mitigate the bias, or fine-tuning the multilingual model can effectively decrease the bias. For all languages, the alignment approach reduces bias and is a better solution for low-resource languages.

Most of the research on bias is limited to English, and our work contributes to studying bias in multiple languages including relatively low-resource languages. Our study shows the variation of ethnic bias across languages with the same set of templates and attributes translated into multiple languages.

Acknowledgement

This demo website is supported by KAIST Institute for Artificial Intelligence.