Show HN: Word2Vec Gender Bias Explorer

chanind.github.io

4 points·chanind··3 comments

Cool project, although it seems to think the word “male” has moderate female bias... word embeddings can be fussy.

I would think you could just compare cosine similarities between the entered word and words like male/female or man/woman (not sure if this is what you’re doing).

Why shouldn't the word "male" have moderate female bias? It depends on the contexts in which it appears across the corpus.

As a biological sex, I can think of numerous ways it will co-occur with "feminine" words - the first few sentences of the Wikipedia entry for "Male" being a good example:

"A male gamete can fuse with a larger female gamete, or ovum, in the process of fertilization. A male cannot reproduce sexually without access to at least one ovum from a female, but some organisms can reproduce both sexually and asexually."

This demo shows gender bias, which as a concept is deep-rooted in most if not all natural languages. Consequently, it may be more accurate to say the word "male" is tips the scale towards "feminine" or "womanly" rather than "masculine" or "manly".

Ugh that's embarrassing. It seems like it works for most words, but has a few obnoxious outliers like "male" that I can't figure out why they score as female. It might just be a quirk of the dataset or something :/

Reply on news.ycombinator.com