Gender Ethnicity Adjective Profession woman man non-binary (person) African-American American Indian Asian Black Caucasian East Asian First Nations Hispanic Indigenous American Latino Latinx Native American Multiracial Pacific Islander South Asian Southeast Asian White Male-leaning: ambitious assertive confident decisive determined intelligent outspoken self-confident stubborn unreasonable committed Female-leaning: supportive sensitive emotional gentle honest modest compassionate considerate pleasant accountant aerospace engineer aide air conditioning installer architect author baker bartender career counselor carpenter carpet installer cashier CEO childcare worker civil engineer claims appraiser cleaner clergy clerk coach community manager compliance officer computer programmer computer support specialist computer systems analyst cook correctional officer courier credit counselor customer service rep. T able 4: A list of the social attributes (gender and ethnicity) and target attributes. All "professions" prompts specify a profession value. Statistically significant results are bolded. "Latinx" is the most frequently appearing ethnicity term and "woman" is the most frequent gender word.

Bias is a disproportionate prejudice in favor of one side against another. Due to the success of transformer-based Masked Language Models (MLMs) and their impact on many NLP tasks, a systematic evaluation of bias in these models is needed more than ever. While many studies have evaluated gender bias in English MLMs, only a few works have been conducted for the task in other languages. This paper proposes a multilingual approach to estimate gender bias in MLMs from 5 languages: Chinese, English, German, Portuguese, and Spanish. Unlike previous work, our approach does not depend on parallel corpora coupled with English to detect gender bias in other languages using multilingual lexicons. Moreover, a novel model-based method is presented to generate sentence pairs for a more robust analysis of gender bias, compared to the traditional lexicon-based method. For each language, both the lexicon-based and model-based methods are applied to create two datasets respectively, which are used to evaluate gender bias in an MLM specifically trained for that language using one existing and 3 new scoring metrics. Our results show that the previous approach is data-sensitive and not stable as it does not remove contextual dependencies irrelevant to gender. In fact, the results often flip when different scoring metrics are used on the same dataset, suggesting that gender bias should be studied on a large dataset using multiple evaluation metrics for best practice.

Societal biases present in pre-trained large language models are a critical issue as these models have been shown to propagate biases in countless downstream applications, rendering them unfair towards specific groups of people. Since large-scale retraining of these models from scratch is both time and compute-expensive, a variety of approaches have been previously proposed that de-bias a pre-trained model. While the majority of current state-of-the-art debiasing methods focus on changes to the training regime, in this paper, we propose data intervention strategies as a powerful yet simple technique to reduce gender bias in pre-trained models. Specifically, we empirically show that by fine-tuning a pre-trained model on only 10 de-biased (intervened) training examples, the tendency to favor any gender is significantly reduced. Since our proposed method only needs a few training examples, our few-shot debiasing approach is highly feasible and practical. Through extensive experimentation, we show that our debiasing technique performs better than competitive state-of-the-art baselines with minimal loss in language modeling ability.

Given messages The elimination of discrimination is an important with the same content for different genders, issue that our society is facing. Learning from dialogue models could produce biased responses, human behaviors, machine learning algorithms which have been measured in terms of their politeness have been proven to inherit the prejudices from and sentiment, as well as the existence of humans (Mehrabi et al., 2019). A variety of AI applications biased words (Liu et al., 2019a). Table 1 shows one have demonstrated common prejudices example from a generative dialogue model trained towards particular groups of people (Rodger and on the Twitter dialogue corpus. When we change Pendharkar, 2004; Howard and Borenstein, 2018; the words in the message from "he" to "she", the responses Rose, 2010; Yao and Huang, 2017; Tolan et al., produced by the dialogue model are quite 2019). It is evident from recent research that different. In particular, the dialogue model generates learning-based dialogue systems also suffer from responses with negative sentiments for females.