Exploring Gender Bias in Word2Vec

tags:: #bias #AI #gender #NLP projects::

• AI (Artificial Intelligence) models can often reflect biases we have as humans
  • Specifically, they reflect biases of the dataset that is fed to train them
  • What biases exist in machine learning models
• In this github page, gender bias in the Word2Vec model is covered
• [Try this bias tool, put in a word and see if there exists a gener bias][https://chanind.github.io/word2vec-gender-bias-explorer/#/query?sentence=librarian]
  • I preloaded the word "Librarian" in the url, this word shows a strong female association
• You can add and subtract words from each other to get new words, like "Tokyo" - "Japan" + "France" ~= "Paris", the Word2Vec atomic note breaks this down more in depth
• We can view a gender dimention by performing calculations like: "Queen" - "woman" + "man" ~= "King"
  • In cases like this the words are the same except the gender they refer to
  • This may be a bit technical but bear with me:
    • The researches create a mapping from male associate words to female associated words (and vice versa)
    • Just imagine these words as numbers, we can find an average number that converts a male associated word (King) to the corresponding female word (Queen)
    • Using these male->female/female->male mappings we can find out if words have a female or male association
• Using these computed gender vectors, we can see some interesting results:
  • There are many words that have female (or male) biases
    • For example, "philosopher" has a strong male assoiation where "nurse" has a strong female association
  • Interestingly, most words have (at least a slight) male bias, even words like "the"/"and", why?
    • As the article suggests, this could be because the Google News dataset that was used to train Word2Vec talk about men more often than women.
  • Some words have unexpeced biases, the word "husband" is apparently a female biased word, why?
    • "wife" is too female biased, after checking, this of course makes sense
    • It's suggested that "husband" is female associated because it is often surrounded by a strongly female word, for example "her husband" or "Julie's husband"

Citational Information¶

Exploring Gender Bias in Word2Vec (2021) chanind.github.io. Available at: https://chanind.github.io/nlp/2021/06/10/word2vec-gender-bias.html (Accessed: 9 February 2023).

Related Links¶

• I find it interesting that we can learn about what biases our society holds by understanding the data deeply.
• For example, the fact that %70 of words inherintly have a male bias :
  • might indicate gender roles and societal norms that lead to men in media
  • This could also show a lack of representation of women in leadership positions
  • sports are male dominated domains

• Careers that are associated with men/women

  
  - This indicates other historical and cultural stereotypes - Systemic gender biases - Representaition of men/women in different professions