On Measuring Social Biases in Sentence Encoders
This addresses the problem of detecting social biases in widely-used sentence encoders for AI fairness, though it is incremental as it builds on prior word-level bias tests.
The authors extended the Word Embedding Association Test to measure social biases in sentence encoders, testing methods like ELMo and BERT for biases such as gender and race, and found mixed results with suspicious sensitivity patterns.
The Word Embedding Association Test shows that GloVe and word2vec word embeddings exhibit human-like implicit biases based on gender, race, and other social constructs (Caliskan et al., 2017). Meanwhile, research on learning reusable text representations has begun to explore sentence-level texts, with some sentence encoders seeing enthusiastic adoption. Accordingly, we extend the Word Embedding Association Test to measure bias in sentence encoders. We then test several sentence encoders, including state-of-the-art methods such as ELMo and BERT, for the social biases studied in prior work and two important biases that are difficult or impossible to test at the word level. We observe mixed results including suspicious patterns of sensitivity that suggest the test's assumptions may not hold in general. We conclude by proposing directions for future work on measuring bias in sentence encoders.