Machine Learning and Friends Lunch (Online)

Title: Addressing Biases in Pre-Trained Language Models

Abstract: Large-scale pre-trained Language Models (LMs) have seen enormous success across many NLP tasks. However, there is evidence that these models reflect societal biases in the learned representations and the downstream applications. We find that de-biasing approaches in the contextual embeddings space can be ineffective at the downstream task level, since biases can be re-introduced during fine-tuning. We investigate whether biases internalized by large LMs during pre-training affect downstream behavior after fine-tuning. For two classification tasks, we find that reducing representation bias with interventions before fine-tuning has little impact on task-specific predictions (after fine-tuning). In fact, downstream disparities are better explained by biases in the fine-tuning dataset. Motivated by these observations, we present a de-biasing approach based on stochastic word dropout. Our approach acts on the task-specific data during fine-tuning and it selectively attenuates contribution from words which are highly correlated with words indicative of societal biases. Our approach encourages practitioners to focus more on the task-specific dataset and the context-specific harms.

Bio: Swetasudha (Sweta) Panda is a research scientist at the Machine Learning Research Group of Oracle Labs located in Burlington, MA. Previously, she received a Ph.D. in Computer Science from Vanderbilt University, working with Yevgeniy Vorobeychik. She graduated with B.Tech. in Electrical Engineering from Indian Institute of Technology, Kharagpur. Her research interests span fairness-aware ML and NLP, algorithms for social good, stochastic planning and computational game theory.

To obtain the Zoom link for this event, please see the event announcements from MLFL on the college email lists or contact Wenlong Zhao.