Bias and Discrimination in Natural Language Processing
Natural Language Processing (NLP) is the ability that helps computer program understand, interpret and manipulate human language. In addition, it is the way of abstracting the meaningful use of language in situations. Computers look beyond individual words or phrases and understand the context they are delivered in. The applications such as Alexa and Siri understand when you ask them to perform certain tasks. As the NLP is rising and has shown success in building many applications, they have bias and discrimination in them. The bias in NLP is complex and exacerbate concern. Several biases currently exist in our modern NLP systems such as demographic bias and gender bias just to name a few. However, most people find it difficult to believe that Artificial Intelligence can be biased.
NLP models can pick up a variety of bias and discrimination based on cultural associations and adverse social biases. Bias and discrimination commonly exist across all languages. Models such as word embedding have identified to have gender bias and addressed by measuring projections onto linear gender-related subspaces of word representation (Huang, et al., 2020). Word embedding is a popular framework for machine learning in NLP. It is more gender bias because many people still talk and write with these biases. Gender bias exists strongly in the set of occupation (Sosnick, 2017) such as the word doctor is closer to the word ‘he’ than ‘she’. Sentiment analysis found to have systematic biases such as race and gender across more than 200 systems (Huang, et al., 2020). Similarly, social discrimination (Papakyriakopoulos, et al., 2020) is another common problem in NLP, especially in word embedding. It refers to the discrimination developing from members of one social group towards members of another. Some of these papers also discuss other types of biases such as technical bias and emergent bias (Papakyriakopoulos, et al., 2020).
Many papers suggest different strategies for addressing bias in NLP. The bias can be reduced by resampling training data from a larger dataset (Diaz, et al., 2018). This method reduced the age-related bias. Rather than debiasing word embeddings, Huang and friends suggest data augmentation as a remedy to occupation-based gender biases (Huang, et al., 2020). Similarly, embedding regularization and sentiment regularization helps to reduce the sentiment biases without losing their semantics. The tools such as distributive justice that serves to identify what is wrong with gender bias, ethics of care and AI technologies as moral are suggested to overcome gender bias (Wellner, 2020). The ENDB method is used to align the embeddings to ENDEB space. The researchers made the training corpus gender-balanced for every occupation by upsampling to make the model free of corpus bias (Zhao, et al., 2020). The bias detection frameworks (Cirilo, et al., 2020) for fairness can be used to achieve fairness in the NLP. One method is to remove some sensitive and personal information such as gender and all other possible features during the data processing. Also, the use of visualization, logical statements and dimensionality reduction techniques can be used to achieve interoperability to achieve interpretable models (Cirilo, et al., 2020). Human judgement is still required to ensure NLP is fair while making decisions (Silbberg & Manyika, 2019).
Many papers have a common bias which is gender bias in NLP. Also, every paper believes that bias can be introduced during the data collection as the way they are collected and selected. Similarly, most of the paper suggest that oversampling the train data helps to minimize the bias and discrimination in the model. Almost every paper believe that bias and discrimination can be reduced but not completely remove from the model as there are way too many biases that can exist in data and it is almost impossible to detect them. Also, they believe that all these biases cannot be solved using mathematical intuitions. Some paper (Wellner, 2020) has found that there is a limitation in other languages corpus, and they had to experiment only on certain languages such as English and French.
This paper systematically compares different types of bias in NLP and their mitigation strategies. Many papers have shown that significant gender bias exists in word embedding and more work is required to explore other types of biases. Several papers propose different ways to minimize a different kind of cultural and social biases. Also, not all types of bias and discrimination can be detected as it is not always possible using formal mathematical techniques because of their complexity. Furthermore, researchers are considering different computational techniques to understand the underrepresented groups such as in sentiment analysis. The social bias in algorithms can be used to understand how unrecognized social bias activates at scale. Most of the research work is done using English corpora and it is less known about the bias in multilingual embedding and many researchers will be doing more research and experiment across the different corpus.
References
Cirilo, D. et al., 2020. Sex and gender differences and biases in artificial intelligence for biomedicine and healthcare. [Online]
Available at: https://www.nature.com/articles/s41746-020-0288-5
[Accessed 2 August 2020].Diaz, M., Johnson, I., Piper, A. M. & Gergle, D., 2018. Addressing Age-Related Bias in Sentiment Analysis. [Online]
Available at: https://dl.acm.org/doi/pdf/10.1145/3173574.3173986
[Accessed 1 August 2020].Ghosh, D., 2017. AI is the future of hiring, but it’s far from immune to bias. [Online]
Available at: https://scholar.harvard.edu/files/dipayan/files/ai_in_hiring_can_lead_to_algorithmic_bias_-_quartz_at_work.pdf
[Accessed 2 August 2020].Huang, P.-S.et al., 2020. Reducing Sentiment Bias in Language Models via Counterfactual Evaluation. [Online]
Available at: https://arxiv.org/pdf/1911.03064.pdf
[Accessed 1 August 2020].Lloyd, K., 2019. Bias Amplification in Artificial Intelligence Systems. [Online]
Available at: https://arxiv.org/pdf/1809.07842.pdf
[Accessed 2 August 2020].Papakyriakopoulos, O., Hegelich, S., Marco, F. & Serrano, J. C. M., 2020. Bias in Word Embeddings. [Online]
Available at: https://dl.acm.org/doi/pdf/10.1145/3351095.3372843
[Accessed 1 August 2020].Silbberg, J. & Manyika, J., 2019. Notes from the AI frontier: Tackling bias in AI (and in humans). [Online]
Available at: https://www.mckinsey.com/~/media/McKinsey/Featured%20Insights/Artificial%20Intelligence/Tackling%20bias%20in%20artificial%20intelligence%20and%20in%20humans/MGI-Tackling-bias-in-AI-June-2019.pdf
[Accessed 2 August 2020].Sosnick, M. A., 2017. Exploring Fairness and Bias in Algorithms and Word Embedding. [Online]
Available at: https://fisher.wharton.upenn.edu/wp-content/uploads/2019/06/Thesisi_Sosnick.pdf
[Accessed 1 August 2020].Wellner, G. P., 2020. When AI is Gender-biased: The Effects of Biased AI on the Everyday Experiences of Women. [Online]
Available at: http://www.humanamente.eu/index.php/HM/article/view/307/273
[Accessed 2 August 2020].Zhao, J. et al., 2020. Gender Bias in Multilingual Embeddings. [Online]
Available at: https://www.researchgate.net/profile/Subhabrata_Mukherjee/publication/340660062_Gender_Bias_in_Multilingual_Embeddings/links/5e97428692851c2f52a6200a/Gender-Bias-in-Multilingual-Embeddings.pdf
[Accessed 2 August 2020].
