Measuring Social Biases in Grounded Vision and Language Embeddings

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 998–1008 June 6–11, 2021. ©2021 Association for Computational Linguistics 998 Measuring Social Biases in Grounded Vision and Language Embeddings Candace Ross, Boris Katz & Andrei Barbu CSAIL, Massachusetts Institute of Technology {ccross,boris,abarbu}@mit.edu Abstract We generalize the notion of measuring so- cial biases in word embeddings to visually grounded word embeddings. Biases are present in grounded embeddings, and indeed seem to be equally or more significant than for un- grounded embeddings. This is despite the fact that vision and language can suffer from dif- ferent biases, which one might hope could attenuate the biases in both. Multiple ways exist to generalize metrics measuring bias in word embeddings to this new setting. We intro- duce the space of generalizations (Grounded- WEAT and Grounded-SEAT) and demonstrate that three generalizations answer different yet important questions about how biases, lan- guage, and vision interact. These metrics are used on a new dataset, the first for grounded bias, created by augmenting standard linguis- tic bias benchmarks with 10,228 images from COCO, Conceptual Captions, and Google Im- ages. Dataset construction is challenging be- cause vision datasets are themselves very bi- ased. The presence of these biases in systems will begin to have real-world consequences as they are deployed, making carefully measuring bias and then mitigating it critical to building a fair society. 1 Introduction Since the introduction of the Implicit Association Test (IAT) by Greenwald et al. (1998), we have had the ability to measure biases in humans. Many IAT tests focus on social biases, such as inher- ent beliefs about someone based on their racial or gender identity. Social biases have negative im- plications for the most marginalized people, e.g., applicants perceived to be Black based on their names are less likely to receive job interview call- backs than their white counterparts (Bertrand and Mullainathan, 2004). Caliskan et al. (2017) introduce an equivalent of the IAT for word embeddings, called the Word Embedding Association Test (WEAT), to measure word associations between concepts. The results of testing bias in word embeddings using WEAT par- allel those seen when testing humans: both reveal many of the same biases with similar significance. May et al. (2019) extend this work with a met- ric called the Sentence Encoder Association Test (SEAT), that probes biases in embeddings of sen- tences instead of just words. We take the next step and demonstrate how to test visually grounded em- beddings, specifically embeddings from visually- grounded BERT-based models by extending prior work into what we term Grounded-WEAT and Grounded-SEAT. The models we evaluate are ViL- BERT (Lu et al., 2019), VisualBERT (Li et al., 2019) , LXMert (Tan and Bansal, 2019) and VL- BERT (Su et al., 2019). Grounded embeddings are used for many conse- quential tasks in natural language processing, like visual dialog (Murahari et al., 2019) and visual question answering (Hu et al., 2019). Many real- world tasks such as scanning documents …

Content truncated to 500 words