Recognizing and Reducing Bias
                 in NLP Applications
                Dirk Hovy Università Bocconi, Italia
                    dirk.hovy@unibocconi.it

                                Abstract
As NLP technology becomes used in ever more settings, it has ever more
impact on the lives of people all around the world. As NLP practitioners, we
have become increasingly aware that we have the responsibility to evaluate
the effects of our research and prevent or at least mitigate harmful outcomes.
This is true for academic researchers, government labs, and industry devel-
opers. However, without experience of how to recognize and engage with
the many ethical conundrums in NLP, it is easy to become overwhelmed and
remain inactive. One of the most central ethical issues in NLP is the impact
of hidden biases that affect performance unevenly, and thereby disadvantage
certain user groups.
    This tutorial aims to empower NLP practitioners with the tools spot these
biases, and a number of other common ethical pitfalls of our practice. We will
cover both high-level strategies, as well as go through specific case sample
exercises. This is a highly interactive workshop with room for debate and
questions from the attendees. The workshop will cover the following broad
topics:
    • Biases: Understanding the different ways in which biases affect NLP
      data, models, and input representations, including including strategies
      to test for and reduce bias in all of them.
   • Dual Use: Learning to anticipate how a system could be repurposed
     for harmful or negative purposes, rather than its intended goal.
   • Privacy: Protecting the privacy of users both in corpus construction
     and model building.