With the starting point that implicit human biases are reflected in the statistical regularities of language, it is possible to measure biases in static word embeddings. With recent advances in natural language processing, state-of-the-art neural language models generate dynamic word embeddings dependent on the context in which the word appears. Current methods of measuring social and intersectional biases in these contextualized word embeddings rely on the effect magnitudes of bias in a small set of pre-defined sentence templates. We propose a new comprehensive method, Contextualized Embedding Association Test (CEAT), based on the distribution of 10,000 pooled effect magnitudes of bias in embedding variations and a random-effects model, dispensing with templates. Experiments on social and intersectional biases show that CEAT finds evidence of all tested biases and provides comprehensive information on the variability of effect magnitudes of the same bias in different contexts. Furthermore, we develop two methods, Intersectional Bias Detection (IBD) and Emergent Intersectional Bias Detection (EIBD), to automatically identify the intersectional biases and emergent intersectional biases from static word embeddings in addition to measuring them in contextualized word embeddings. We present the first algorithmic bias detection findings on how intersectional group members are associated with unique emergent biases that do not overlap with the biases of their constituent minority identities. IBD achieves an accuracy of 81.6% and 82.7%, respectively, when detecting the intersectional biases of African American females and Mexican American females. EIBD reaches an accuracy of 84.7% and 65.3%, respectively, when detecting the emergent intersectional biases unique to African American females and Mexican American females (random correct identification probability ranges from 1.0% to 25.5%).
Algorithmic bias is the systematic preferential or discriminatory treatment of a group of people by an artificial intelligence system. In this work we develop a random-effects based metric for the analysis of social bias in supervised machine learning prediction models where model outputs depend on U.S. locations. We define a methodology for using U.S. Census data to measure social bias on user attributes legally protected against discrimination, such as ethnicity, sex, and religion, also known as protected attributes. We evaluate our method on the Strategic Subject List (SSL) gun-violence prediction dataset, where we have access to both U.S. Census data as well as ground truth protected attributes for 224,235 individuals in Chicago being assessed for participation in future gun-violence incidents. Our results indicate that quantifying social bias using U.S. Census data provides a valid approach to auditing a supervised algorithmic decision-making system. Using our methodology, we then quantify the potential social biases of 100 million ridehailing samples in the city of Chicago. This work is the first large-scale fairness analysis of the dynamic pricing algorithms used by ridehailing applications. An analysis of Chicago ridehailing samples in conjunction with American Community Survey data indicates possible disparate impact due to social bias based on age, house pricing, education, and ethnicity in the dynamic fare pricing models used by ridehailing applications, with effect-sizes of 0.74, 0.70, 0.34, and -0.31 (using Cohen's d) for each demographic respectively. Further, our methodology provides a principled approach to quantifying algorithmic bias on datasets where protected attributes are unavailable, given that U.S. geolocations and algorithmic decisions are provided.
Word embeddings learn implicit biases from linguistic regularities captured by word co-occurrence information. As a result, statistical methods can detect and quantify social biases as well as widely shared associations imbibed by the corpus the word embeddings are trained on. By extending methods that quantify human-like biases in word embeddings, we introduce ValNorm, a new word embedding intrinsic evaluation task, and the first unsupervised method that estimates the affective meaning of valence in words with high accuracy. The correlation between human scores of valence for 399 words collected to establish pleasantness norms in English and ValNorm scores is r=0.88. These 399 words, obtained from social psychology literature, are used to measure biases that are non-discriminatory among social groups. We hypothesize that the valence associations for these words are widely shared across languages and consistent over time. We estimate valence associations of these words using word embeddings from six languages representing various language structures and from historical text covering 200 years. Our method achieves consistently high accuracy, suggesting that the valence associations for these words are widely shared. In contrast, we measure gender stereotypes using the same set of word embeddings and find that social biases vary across languages. Our results signal that valence associations of this word set represent widely shared associations and consequently an intrinsic quality of words.
The recent COVID-19 pandemic, which was first detected in Wuhan, China, has been linked to increased anti-Chinese sentiment in the United States. Recently, Broniatowski et al. found that foreign powers, and especially Russia, were implicated in information operations using public health crises to promote discord -- including racial conflict -- in American society (Broniatowski, 2018). This brief considers the problem of automatically detecting changes in overall attitudes, that may be associated with emerging information operations, via artificial intelligence. Accurate analysis of these emerging topics usually requires laborious, manual analysis by experts to annotate millions of tweets to identify biases in new topics. We introduce extensions of the Word Embedding Association Test from Caliskan et. al to a new domain (Caliskan, 2017). This practical and unsupervised method is applied to quantify biases being promoted in information operations. Analyzing historical information operations from Russia's interference in the 2016 U.S. presidential elections, we quantify biased attitudes for presidential candidates, and sentiment toward Muslim groups. We next apply this method to a corpus of tweets containing anti-Chinese hashtags. We find that roughly 1% of tweets in our corpus reference Russian-funded news sources and use anti-Chinese hashtags and, beyond the expected anti-Chinese attitudes, we find that this corpus as a whole contains pro-Russian attitudes, which are not present in a control Twitter corpus containing general tweets. Additionally, 4% of the users in this corpus were suspended within a week. These findings may indicate the presence of abusive account activity associated with rapid changes in attitudes around the COVID-19 public health crisis, suggesting potential information operations.
We seek to determine whether state-of-the-art, black box face recognition techniques can learn first-impression appearance bias from human annotations. With FaceNet, a popular face recognition architecture, we train a transfer learning model on human subjects' first impressions of personality traits in other faces. We measure the extent to which this appearance bias is embedded and benchmark learning performance for six different perceived traits. In particular, we find that our model is better at judging a person's dominance based on their face than other traits like trustworthiness or likeability, even for emotionally neutral faces. We also find that our model tends to predict emotions for deliberately manipulated faces with higher accuracy than for randomly generated faces, just like a human subject. Our results lend insight into the manner in which appearance biases may be propagated by standard face recognition models.
Artificial intelligence and machine learning are in a period of astounding growth. However, there are concerns that these technologies may be used, either with or without intention, to perpetuate the prejudice and unfairness that unfortunately characterizes many human institutions. Here we show for the first time that human-like semantic biases result from the application of standard machine learning to ordinary language---the same sort of language humans are exposed to every day. We replicate a spectrum of standard human biases as exposed by the Implicit Association Test and other well-known psychological studies. We replicate these using a widely used, purely statistical machine-learning model---namely, the GloVe word embedding---trained on a corpus of text from the Web. Our results indicate that language itself contains recoverable and accurate imprints of our historic biases, whether these are morally neutral as towards insects or flowers, problematic as towards race or gender, or even simply veridical, reflecting the {\em status quo} for the distribution of gender with respect to careers or first names. These regularities are captured by machine learning along with the rest of semantics. In addition to our empirical findings concerning language, we also contribute new methods for evaluating bias in text, the Word Embedding Association Test (WEAT) and the Word Embedding Factual Association Test (WEFAT). Our results have implications not only for AI and machine learning, but also for the fields of psychology, sociology, and human ethics, since they raise the possibility that mere exposure to everyday language can account for the biases we replicate here.