IQOQO Uncovers Gender Bias in Media using NLP
Women comprise 51% of the population, but when it comes to public life they are grossly underrepresented. Out of the 435 members of the United States Congress, only 19.4% are women. Sadly, the numbers are much worse in the business world. Only 24 Fortune 500 companies have women as CEOs. Looking at the Women in S&P 500 Companies pyramid, we can see that while women are 45% of employees at the base of the pyramid, that percentage drops the higher up you go up, only reaching the 5% mark for CEO roles.
Representation of women in politics and business may be lacking today, but we should understand a reality where women who occupy lower-ranking positions will become tomorrow’s leaders. In order to get there, however, we need to start advancing equality in everyday life.
And what is more present in “everyday life” than the media?
We decided to see if women were as underrepresented in the media as they are in politics and business. To do this, we used the work of Neal Caren as our project’s baseline. He used Natural Language Processing (NLP) on a week’s worth of New York Time publications to assess female representation. We augmented this work using the All the news dataset which is currently a collection of 143,000 news items across 15 media outlets.
We know that simply looking at the number of mentions is not the best indicator for how women are represented, but it is a starting point. In a future post, we hope to improve on this experiment by adding in a contextual analysis to get a clearer picture of how women are being spoken about.
To make the job run faster, we parallelized execution of the job in the cloud using IQOQO (disclosure, I am IQOQO’s CTO).
With IQOQO, I was able to develop the code on my machine with a minimal sample dataset. Once my code was ready for execution all I had to do was upload the code and data to IQOQO and let IQOQO dispatch the data and run the analysis in parallel. By distributing the job, I freed up my computer which gave me the time, and processing power, to write this blog.
The total runtime of the test was about 3 minutes per 250MB data file. (By using IQOQO, I can run this test periodically and continue to compare results; IQOQO creates a time capsule of jobs run in the cloud.) Also, I simply like using IQOQO, as stated before, I am IQOQO’s CTO, what other tool would I use?
All in all, we reviewed 143,000 news items from fifteen publications. The number of articles from each publication varied between about 5,000 per publication to almost 25,000. We used this data to construct two results. A female-to-male sentences ratio (excluding any gender-neutral sentences) and a list of words most typically associated with males and females words (the list can be viewed here — words marked in red are words that we would like to see more frequently associated with females the next time we run the test).
We were amazed to learn that many management and business-oriented words were used eight times more often in male-oriented sentences than female sentences. There are 1,001 appearances of the word “executive” in male sentences and only 181 in female sentences, 942 instances of the word “leader” in male sentences compared to only 130 for females, while “candidate” appeared 1,268 times for males and 361 times for females. For one of my favorite words, “money”, it appeared in 977 sentences about men and only 294 sentences about women.
Sheryl Sandberg once said, “We can’t become what we don’t see.” To see more equality between women and men, we must have more conversations about women and the roles they occupy in our lives. By examining these 143,000 articles, we learned that we still have a long road ahead.
If you run multiple NLP experiments and want to see how IQOQO can help you get results faster, contact me at iqoqomp[at]iqoqo.co and take IQOQO for a spin.
