Abstract

With the world striving to be an equal opportunity place, instances of social biases are like a step backward. They are targeted towards individuals or groups based on their social identities like race, religion, gender, caste, and so on. The sources of these social biases are mainly humans and AI-assisted platforms like social media, search engines, conversational AI, etc. which are being used by humans for quick decision making. Thus, it is essential to identify and mitigate the social biases coming from AI algorithms to control their impact. We plan to cover the nuances of this issue through this tutorial. The tutorial aims to cover the fundamentals of social bias, including their types, sources and effects in different NLP tasks. Some recent work in the domain of bias detection and mitigation with the state of the art will also be discussed. The tutorial then highlights the challenges yet to be addressed by the community and introduces the specific problem of social biases in Hollywood and Bollywood movie scripts. We will focus on the importance of the relevant contexts in social bias detection and share our experiences of biased text extraction from related domains. The tutorial will also discuss the application of explainability tools for bias detection tasks and some hands-on demonstrations.

Tutorial Prerequisite

Attendees should be familiar with the fundamentals of linear algebra and dimensionality reduction. Basic NLP knowledge would be beneficial but is not required. The audience is not assumed to be familiar with bias in NLP; however, those who are will find the tutorial useful.