# On Measuring Social Biases in Sentence Encoders

## TL;DR

This paper extends bias measurement methods from word embeddings to sentence encoders, evaluating models like ELMo and BERT for social biases, revealing complex patterns and suggesting future research directions.

## Contribution

It introduces a method to measure social biases in sentence encoders and applies it to state-of-the-art models, highlighting challenges and areas for further investigation.

## Key findings

- Sentence encoders exhibit social biases similar to word embeddings.
- Some biases are difficult or impossible to test at the sentence level.
- The bias measurement test may have limitations due to underlying assumptions.

## Abstract

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.