For years, Google Translate has been somewhat biased in recognizing gender by connecting positions work with a specific gender: a doctor with the male gender and a nurse with the female gender.
Yesterday, Google's AI team released a new model that will probably solve the problem. The new AI is already available but provides support for gender translations from English to Spanish and from Finnish, Hungarian and Persian to English.
In a blog post, the company explained that the its old approach gave too many errors in the final results.
The team also reported that although it used a Neural Machine Translation (NMT), the system could not translate gender for more than 40% of the queries.
Left: An early example of translating a gender-neutral English phrase into a Spanish gender equivalent. In this case, only one example can be given that can be discriminated against. Right: The new Translation provides a choice of (female - male) gender.
To train the AI, the team used for the first time millions sentences in the English language to create a basic model.
Then, the researchers they trained it for her appearance of the female gender in a request that requires the male gender and vice versa. Finally, the results from both directions were merged to create a model that gives two options.