Measuring Implicit Human Biases
Through the Statistical Properties of Language

 

Racialization

 

ex. Thug

Happens when a word with no preexisting racial connotation comes to describe people of color asymmetrically.

 

Micro-aggressions

 

ex. Articulate

This word, specifically, works as a micro-aggression

In the US, standard linguistic varieties are ideologically associated with Whiteness and formal education

Calling a minority articulate notes surprisal at standard language usage in the absence of these traits

 

Covert Racism in Linguistics

 

Racialization occurs when regular words, with no racial connotation, become asymmetrically applied to the descriptions of actions or events of actors of color.

Covert Racialization in Sports Journalism

  • The Myth of Unbeatable Black Athleticism

    Folk ideology, mistakenly applied throughout US history

    Black athletes are not exceptional or naturally suited for physical activity, or for violent displays or prowess in team or individual sport

  • Black Athletes in Sports Journalism

    Black athletes are described predictably differently as Exceptionality and Animalistic traits dominating the conversation

  • White Athletes in Sports Journalism

    White athletes are described in predominately Leadership or Skill-based terms

The Dataset and The Procedure

Racialized SEmantics in Athletics Corpus (RSEAC)

  • 120 Athletes

  •  60 White; 60 Black

  • 30 Male; 30 Female

  • 15,500 lexemes

  • 8.5 million total words

    Combines previous methodologies

    • Metadata behind Sports Journalism

    • Highly controlled journalistic frame

    • Longform and event pieces, describing individuals, not teams

  • Variation based on patterns of actual disparities baked into the data, making racialization isolable.

Latent Semantic Analysis

Using Word Vector Models trained on the Google corpus, human participant’s results are recreated from Implicit Association Tasks.

Check Out The Procedure

N= Number of Subjects NT= Target Words NA= Attribute Words D = Effects Sizes P = P values

Word Vectors created from Target word stimuli and Attribute word stimuli

Collection Methods

1. Search Engine Optimization Search

➢ Search Engine Results Pages 

2. Advanced Search

➢ “Serena Williams”
3. Cleaning and concatenation bash script

➢  Brings down text per article, removes noise

➢  Saves as Subcorpora

Machine Learning

Machine learning is a means to derive artificial intelligence by discovering patterns in existing data. Here, we show that applying machine learning to ordinary human language results in human-like semantic biases.

“We replicated a spectrum of known biases, as measured by the Implicit Association Test, using a widely used, purely statistical machine-learning model trained on a standard corpus of text from the World Wide Web. Our results indicate that text corpora contain recoverable and accurate imprints of our historic biases, whether morally neutral as toward insects or flowers, problematic as toward race or gender, or even simply veridical, reflecting the status quo distribution of gender with respect to careers or first names. Our methods hold promise for identifying and addressing sources of bias in culture, including technology (Caliskan et al. 2018).

  • Support Vector Machine—Counting Stuff

  • Random Forest Modeling Task—Analyzing Counted Stuff

Support Vector Machine

In machine learning, support-vector machines are supervised learning models with associated learning algorithms that analyze data for classification and regression analysis.

  • This Learning Algorithm is trained to predict athlete race from lexical token counts

  •  In this study, the subcorpora is organized by athlete

Takeaways

The word Culture occurs in a 2:1 ratio

  • Black athletes are discussed as

    • Infusing their own culture into the sport

    • Altering the sport’s culture with their ethnic presence

  • Culture occurrences in the White subcorpus

    •  The culture of the sport itself, not White culture as such

    • Not in reference to the athlete

  • Two very different lexical senses of culture, the usage of which is determined by the race of the referent

The word Tactics occurs in a 43:43 ratio

  • The ratio of 43:43 reveals that the word tactics was the only monitored word that showed an equality in usage among the Black and White Subcorpus.

The word Rolex occurs in a 11:275 ratio

  • The ratio of 11:275 reveals that the word Rolex is heavily racialized, as it was much more often associated with the White athletes than the Black athletes in the sample of mainstream sports broadcasting.

  • In addition, it is fascinating to see a variable of class emerge as this word is so disproportionately associated with White athletes.

Conclusions

The findings from this Learning Algorithm are not only significant, but staggering. The data reveals that there are racial biases rooted in the language used in mainstream sportscasting. The use of racialized language by broadcasters representing dominant sports networks has a negative trickle-down effect. If the information received by audiences is littered with tones of racism and microagressions, then a culture of racialized language will surely proliferate. Broadcasters and reporters in positions with access to large public platforms have a duty to communicate equitably. The words and practices that they disseminate reinforce a rhetoric acceptable to be used by the general public. In order to effectively eradicate and disable institutionalized racism in sport, a sociolinguistic lens must be applied to target these instances.