ariboyarsky
commented
7 years ago Additionally, a further question that may be of interest is the gender bias of specific topics. For instance, if we cluster job listings by topics do some topics score higher in gender bias? Can we then use topic models to inform our estimates of gender bias?
I'm interested in what the exact method that Texito and Unitive are using to annotate keywords?
Additionally, the methodology to count a weighted total of these keywords seems a little naive. I think the the replacement method is interesting, but I wonder if it allows you truly capture the context of the text. A paper by Dan Jurafsky (https://web.stanford.edu/~jurafsky/pubs/neutrality.pdf) goes into the details of detecting biased language through looking at framing bias and epistemological bias ( i.e. the difference between "murder" and "kill"). It seems that the technique used here may capture epistemological bias (man vs guy). However, will it identify framing bias, i.e. how a sentence may be phrased as an assertion vs. a statement. Then, how may that modify the bias of a sentence, and hence modify your overall score?
Do you think this would be a feasible method for further work? Would you see notable differences in results?