Connotations of male and female jobs

[Richard]

I've been exploring word2vec - a database of words categorised on 50 or so categories all generated from machine learning analysis of massive corpora. I found a fun website where you can conduct your own analyses of the word2vec database: https://lamyiowce.github.io/word2viz/ This gives some default choices to investigate (such as jobs by gender), but also allows you to set up your own analyses. To investigate sexism in language use, first I chose 'Empty' for 'What do you want to see?' Then I set the X-axis as 'good' and 'bad' and the Y-axis as 'he' and 'she'. I then put in pairs of words representing stereotypical male and female jobs (actor - actress, waiter - waitress, priest - priestess, monk - nun, manager - receptionist, pimp - prostitute). In all of these cases, as we would expect the stereotypically female job was higher in the graph (more associated with 'she'). What I found really worrying is that also all of the stereotypically female jobs were to the right (more associated with 'bad'). The only pair I could find where the female job was 'better' than the male is doctor - nurse.

From a critical perspective, the argument that positions of power are more likely to be associated with males than females (e.g. manager - receptionist) has been made numerous times, but I'm not aware of any research suggesting that female jobs are viewed more pejoratively than male jobs. If you put in other pairs, does this pattern hold? Is it worth doing research on this?

Any ideas of other fun ways of exploiting this tool?

[sgtowns]

So for my "exploitation", I did a similar setup to Richard, with a good-bad axis and a male-female axis. I first plotted systems of government. By quadrants, the strongest results were:

* Good Female: democracy
* Good Male: socialism
* Bad Female: (none)
* Bad Male: dictatorship, communism

Then I put in words that might correlate to these different systems of government. The good-bad divide was very stark:

* Good: cooperation, collaboration, sharing, rights, freedom, equality, egalitarian, competition
* Bad: bribe, coup, corruption, exploitation, colluding

So, based on this experience, I would say that 1) the corpus is very biased with regards to the "goodness" of government types and 2) it was easy to cherry-pick words that fit my own biases and ignore words that didn't, so any research using this tool should consider this issue.

But Word2Vec is a very interesting idea that might be useful for language research. For example, based on what I said above, I wonder if it could be used to find "implicit biases" in a corpus.