Analysis of images displayed on google search –climate change
Gender Distributions in Google Images Search Results ... · Web Scraping Download the first 500...
1
Abstract Methods Results Search Term Selection First, 104 professions and 62 college majors were selected from census data [3,4]. To measure people’s beliefs, 23 and 25 participants, respectively, estimated the gender distribution for each profession (e.g. “Percentage of accountants who are men”) and college major (e.g. “Percentage of psychology college students who are men”). Second, 12 alcoholic drinks, 17 music genres and instruments, 15 personality traits, and 20 sports were selected. Note that these domains are strongly gendered, but contrary to the previous ones there is no easily ascertainable ground truth about gender distributions. Processing Google Images Search Results We calculate the proportion of women in the Google Images search results for each of these search terms using the algorithm described below. Additionally, the faces in these results were averaged to form a composite face. Face Detection Support Vector Machine trained on Histogram of Oriented Gradient features Gender Recognition Convolutional Neural Network Web Scraping Download the first 500 Google Images search results Discovering what people really believe is a complex task, especially in contentious domains like gender. In psychology, beliefs are typically assessed using explicit questions or more indirect, implicit measures. However, advances in machine learning coupled with the availability of big data have surfaced new methods for evaluating beliefs. For instance, word embeddings that capture semantic relationships between words have been shown to reflect human-like biases [1,2]. We capitalized on these advances by scraping the results for more than 200 search terms on Google Images, over 100,000 images in total, and comparing the gender distribution for each search term with people’s beliefs and actual statistics. With this approach, we show that internet images reflect beliefs and statistics about gender. Discussion These results indicate that the percentage of women in different professions and college majors corresponds both with census data and participants’ beliefs. Similarly, the images in the domains for which gender distributions are unknown reproduced people’s commonly held beliefs. Additionally, face averaging further supported these findings. In summary, our findings suggest that search results from Google Images – a free, vastly rich, and ubiquitous source of information for millions of people – reflect both beliefs about the world and actual statistics. However, these gender distributions may undercut people’s desires for equal representation between men and women, in which case a commonly used source of information may perpetuate gender inequalities. 1. Professions Security guard Chief executive Flight attendant Model 2. College Majors Person Drinking Beer Philosophy student Psychology student Social work student 3. Alcoholic Drinks 4. Personality Traits 5. Music Genres and Instruments 6. Sports Industrial production technologies student Person Drinking Cosmopolitan Hip Hop Fan Sensitive Person Intelligent Person K-pop Fan Person Bodybuilding Person Cheerleading Gender Distributions in Google Images Search Results: Beliefs and Statistics Juan López Martín 1,2 , Jack Cao 3 , and Mahzarin R. Banaji 3 1 Oxford Internet Institute, University of Oxford; 2 Department of Psychology, Autonomous University of Madrid; 3 Department of Psychology, Harvard University [1] Caliskan et al. (2017). Science; [2] Garg et al. (2017). PNAS; [3] Bureau of Labor Statistics. (2018). CPS; [4] Casselman. (2014). FiveThirtyEight
Transcript of Gender Distributions in Google Images Search Results ... · Web Scraping Download the first 500...
Abstract
Methods
Results
Search Term Selection
First, 104 professions and 62 college majors were selected from censusdata [3,4]. To measure people’s beliefs, 23 and 25 participants,respectively, estimated the gender distribution for each profession (e.g.“Percentage of accountants who are men”) and college major (e.g.“Percentage of psychology college students who are men”).
Second, 12 alcoholic drinks, 17 music genres and instruments, 15personality traits, and 20 sports were selected. Note that these domainsare strongly gendered, but contrary to the previous ones there is no easilyascertainable ground truth about gender distributions.
Processing Google Images Search Results
We calculate the proportion of women in the Google Images searchresults for each of these search terms using the algorithm describedbelow. Additionally, the faces in these results were averaged to form acomposite face.
Face DetectionSupport Vector Machine trained on
Histogram of Oriented Gradient features
Gender RecognitionConvolutional Neural Network
Web ScrapingDownload the first 500 Google
Images search results
Discovering what people really believe is a complex task, especially incontentious domains like gender. In psychology, beliefs are typicallyassessed using explicit questions or more indirect, implicit measures.However, advances in machine learning coupled with the availability ofbig data have surfaced new methods for evaluating beliefs. For instance,word embeddings that capture semantic relationships between wordshave been shown to reflect human-like biases [1,2].
We capitalized on these advances by scraping the results for more than200 search terms on Google Images, over 100,000 images in total, andcomparing the gender distribution for each search term with people’sbeliefs and actual statistics. With this approach, we show that internetimages reflect beliefs and statistics about gender.
DiscussionThese results indicate that the percentage of women in differentprofessions and college majors corresponds both with census data andparticipants’ beliefs. Similarly, the images in the domains for whichgender distributions are unknown reproduced people’s commonly heldbeliefs. Additionally, face averaging further supported these findings.
In summary, our findings suggest that search results from Google Images– a free, vastly rich, and ubiquitous source of information for millions ofpeople – reflect both beliefs about the world and actual statistics.However, these gender distributions may undercut people’s desires forequal representation between men and women, in which case a commonlyused source of information may perpetuate gender inequalities.
1. Professions
Security guard Chief executive Flight attendant Model
2. College Majors
Person Drinking
Beer
Philosophy
student
Psychology
student
Social work
student
3. Alcoholic Drinks
4. Personality Traits
5. Music Genres and Instruments
6. Sports
Industrial production
technologies student
Person Drinking
Cosmopolitan
Hip Hop
Fan
Sensitive
Person
Intelligent
Person
K-pop
Fan
Person
BodybuildingPerson
Cheerleading
Gender Distributions in Google Images Search Results: Beliefs and StatisticsJuan López Martín1,2 , Jack Cao3, and Mahzarin R. Banaji3
1 Oxford Internet Institute, University of Oxford; 2 Department of Psychology, Autonomous University of Madrid;3 Department of Psychology, Harvard University
[1] Caliskan et al. (2017). Science; [2] Garg et al. (2017). PNAS;
[3] Bureau of Labor Statistics. (2018). CPS; [4] Casselman. (2014). FiveThirtyEight
