This talk discusses findings from the Discovering and Attesting Digital Discrimination project at the Department of Informatics, King’s College London. In this project, which focuses on biases in Machine Learning, we proposed a data-driven approach to discover language biases encoded in the vocabulary of discourse communities on social media.
In our approach, we use word embeddings to connect protected attributes to evaluative words found in different language datasets, discovering gender, religion, and ethnic biases. We then used this approach to inform pedagogical settings, including workshops for graduate students at the Department of Digital Humanities and general audiences at the Science Gallery London. The relatively simplistic encoding of words in vector space allows participants to reflect on 1) the ways in which social biases can be perpetuated by ML models, and 2) the epistemic issues of binary categorization that our approach, like many ML-based interventions, integrates.
Location: Remote via Zoom. Link will be sent on the morning of the event.
Date & Time: This talk is a one time event running from 9am to 10am on May 16th, 2022.
Start Time: D-Lab workshops start 10 minutes after the scheduled start time (“Berkeley Time”). We will admit all participants from the waiting room at that time.
Recordings: This D-Lab workshop will be recorded and made available to UC Berkeley participants for a limited time. Your registration for the event indicates your consent to having any images, comments and chat messages included as part of the video recording materials that are made available.