Flammie

Flammie A Pirinen on github pages

Project maintained by flammie Hosted on GitHub Pages — Theme by mattgraham

Flammie on gendering NLP

One of the big annoyances in doing natural language processing especially in Europe is that you have to assign an often binary gender to pretty much everything. This is because some European languages really need to know gender of everything that is talked about. The gender is quite arbitrary designation anyways and in many of the languages it is random for a lot of things but the most problematic part for NLP is probably just tracking genders of humans then, cause that does often match the gender of something else then. So as to summarise the problem in examples:

The three languages have different requirements to mark gender, if you machine translate from one without explicit information it has to be conjured from somewhere. There are few popular approaches to this:

  1. majority votes: he is more common than she, male programmers are more common than female. This is nothing but problems, being wrong slightly less than 50 % of times is not a good thing when mistakes are problematic
  2. gather gendered words in context, this is ok for some few words but gets wrong quite fast

The correct solution for this problem is not to guess when you cannot know. It’s very easy for people to tolerate computer systems that say they too often or *innen too often.

Names with genders

There are traditions in many places to keep names for girls and boys separate, in some countries it is even codified in laws. Problem is, it has never worked even within these countries. If you look a census data or name database, usually best you can get is few odd names with close to 100 % gender distribution and a long row of 90–50%, and that is only within one language community and age group. Crossing some country borders, going 100 years to the past, the statistics get flipped almost more often than not. And that doesn’t even begin to take into account how gender binary has become more outdated as of late.

List of Names

This is just a short collection of names that have broken my machine translations: