Word Embeddings (Also) Encode Human Personality Stereotypes

Proceedings of the Eighth Joint Conference on Lexical and Computational Semantics (*SEM), pages 205–211 Minneapolis, June 6–7, 2019. c⃝2019 Association for Computational Linguistics 205 Word Embeddings (Also) Encode Human Personality Stereotypes Oshin Agarwal University of Pennsylvania oagarwal@seas.upenn.edu Funda Durupinar Bilkent University fundad@cs.bilkent.edu.tr Norman I. Badler University of Pennsylvania badler@seas.upenn.edu Ani Nenkova University of Pennsylvania nenkova@seas.upenn.edu Abstract Word representations trained on text reproduce human implicit bias related to gender, race and age. Methods have been developed to re- move such bias. Here, we present results that show that human stereotypes exist even for much more nuanced judgments such as per- sonality, for a variety of person identities be- yond the typically legally protected attributes and that these are similarly captured in word representations. Specifically, we collected hu- man judgments about a person’s Big Five per- sonality traits formed solely from information about the occupation, nationality or a com- mon noun description of a hypothetical person. Analysis of the data reveals a large number of statistically significant stereotypes in people. We then demonstrate the bias captured in lexi- cal representations is statistically significantly correlated with the documented human bias. Our results, showing bias for a large set of per- son descriptors for such nuanced traits put in doubt the feasibility of broadly and fairly ap- plying debiasing methods and call for the de- velopment of new methods for auditing lan- guage technology systems and resources. 1 Introduction Implicit association tests probe biases individu- als may harbor, by measuring the reaction times of people when asked to sort word stimuli with clearly positive/negative valance and words asso- ciated with racial groups or less morally relevant categories such as insects/flowers and musical in- struments/weapons (Greenwald et al., 1998). Re- cent work has revealed that word representations trained on large text corpora reproduce human bias in preference to flowers and musical instru- ments, but also disturbingly on gender, race and age-related bias (Caliskan et al., 2017). These findings pose a dilemma. Having sys- tems learn that flowers/musical instruments are pleasant and insects/weapons unpleasant appears to be useful common sense knowledge that sys- tems can leverage to better interact with peo- ple1. Having racist, sexist and ageist systems how- ever is highly undesirable, as these are integrated in broader technologies like machine translation, which can reinforce the stereotype2. Stereotypes are highly problematic because even simply evok- ing them can trigger change in behavior (Duguid and Thomas-Hunt, 2015; Spencer et al., 2016). Guided by these compelling arguments, many researchers have started looking for ways to de- bias word representations and language technolo- gies. In response to the examples in the supple- mentary materials in (Caliskan et al., 2017), that Google Translate translates ‘doctor’ as male and ‘nurse’ as female, Google has indeed rolled out a new version of their systems for certain language pairs, in which both translation versions are dis- played3. Similarly, earlier work has zeroed in on the gender bias in word representation and has pro- posed methods for debiasing, which take in a set …

Content truncated to 500 words