Men Also Like Shopping: Reducing Gender Bias Amplification using Corpus-level Constraints

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2979–2989 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Men Also Like Shopping: Reducing Gender Bias Amplification using Corpus-level Constraints Jieyu Zhao§ Tianlu Wang§ Mark Yatskar‡ Vicente Ordonez§ Kai-Wei Chang§ §University of Virginia {jz4fu, tw8cb, vicente, kc2wc}@virginia.edu ‡University of Washington my89@cs.washington.edu Abstract Language is increasingly being used to de- fine rich visual recognition problems with supporting image collections sourced from the web. Structured prediction models are used in these tasks to take advantage of correlations between co-occurring labels and visual input but risk inadvertently en- coding social biases found in web corpora. In this work, we study data and models as- sociated with multilabel object classifica- tion and visual semantic role labeling. We find that (a) datasets for these tasks con- tain significant gender bias and (b) mod- els trained on these datasets further am- plify existing bias. For example, the ac- tivity cooking is over 33% more likely to involve females than males in a train- ing set, and a trained model further ampli- fies the disparity to 68% at test time. We propose to inject corpus-level constraints for calibrating existing structured predic- tion models and design an algorithm based on Lagrangian relaxation for collective in- ference. Our method results in almost no performance loss for the underlying recog- nition task but decreases the magnitude of bias amplification by 47.5% and 40.5% for multilabel classification and visual seman- tic role labeling, respectively. 1 Introduction Visual recognition tasks involving language, such as captioning (Vinyals et al., 2015), visual ques- tion answering (Antol et al., 2015), and visual se- mantic role labeling (Yatskar et al., 2016), have emerged as avenues for expanding the diversity of information that can be recovered from im- ages. These tasks aim at extracting rich seman- tics from images and require large quantities of la- beled data, predominantly retrieved from the web. Methods often combine structured prediction and deep learning to model correlations between la- bels and images to make judgments that otherwise would have weak visual support. For example, in the first image of Figure 1, it is possible to pre- dict a spatula by considering that it is a com- mon tool used for the activity cooking. Yet such methods run the risk of discovering and exploiting societal biases present in the underlying web cor- pora. Without properly quantifying and reducing the reliance on such correlations, broad adoption of these models can have the inadvertent effect of magnifying stereotypes. In this paper, we develop a general framework for quantifying bias and study two concrete tasks, visual semantic role labeling (vSRL) and multil- abel object classification (MLC). In vSRL, we use the imSitu formalism (Yatskar et al., 2016, 2017), where the goal is to predict activities, objects and the roles those objects play within an activity. For MLC, we use MS-COCO (Lin et al., 2014; Chen et al., 2015), a recognition task covering 80 object classes. We use gender bias as …

Content truncated to 500 words