Summary of the results Artificially Correct Hackathon 2021

Illustration, to the left a magnifying glass, to the right a screen with a brain on it.
© Goethe-Institut. Illustration: EL BOUM

Between the 1st and 3rd of October, translators, activists and software developers from all over the world, came together to participate in the Goethe-Institut’s Artificially Correct Hackathon.12 teams had 52 hours (lots of coffee), and one mission – to develop innovative, forward-looking ideas and projects to help tackle bias in language and machine translators.
 

By Alana Cullen and Priyanka Dasgupta

The need to hack

 
Each word has a history, a meaning, a context, and as speakers of language, we have the choice to use these words. Therefore, language is not just words spoken, but a vehicle of social reality. Why say ‘chairman’, as opposed to the more neutral  ‘chair’ or ‘chairperson’? Or why, when thinking of a ‘nurse’, automatically presume the pronoun ‘she’? Stereotypes are embedded not only in mindsets but also words.
 
Consequently, stereotypes also get embedded in artificial intelligence (AI). For instance when translating from Polish to English, Google translate automatically assumes the pronouns for certain tasks when the subject is ambiguous:
 
“She is beautiful”
“He is clever”
“He reads”
“She washes the dishes”
 
We see this bias in AI for machine translations because it is trained off data sets that contain human bias, and it learns the default patterns in our language.
 
Not only does AI have built-in gender bias, but racial too. For example, the term “half blood”, which historically has been used as a racial slur, is now propagated and reinforced based off the popularity of children’s fiction. Consequently, the language bias is perpetuated in technology as AI learns from its maker - and the more prejudice it sees, the more biased the system’s default behaviour will become.
 
Acknowledging this bias gets us one step closer to making more neutral, unbiased machine translation. Of course, true neutrality is not achievable in a world with so many power imbalances, but certainly we can work towards making AI work better for more people. Working towards solutions to this bias, will help us tap into the potential that digital transformation can provide, for everybody.
 

Generating solutions

 
It was these innovative solutions that the participants of the Hackathon focused on. Five challenges or areas of bias in AI were identified, that the teams could ‘hack’ into. These included developing gender-fair post-editing techniques, as well as finding a way to identify sentences that were susceptible to bias in machine translation. 
 
Alongside hacking, participants were also given the opportunity to attend keynote lectures by those working in the field of gender, race and AI studies – Ylva Habel, Dagmar Gromann and Manuel Lardelli, Danielle Saunders and Sheila Beladinejad, who shared their ideas and experience:
 

A complication for translation bias is language change. Our ideas of gender and gendered language and stereotypes change over time.

Danielle Saunders, Hackathon keynote speaker



The hackers were also provided with access to experts in the AI field, who could support them in generating their solutions.
 

I had a fun time attending the Hackathon and putting my expertise for creative solutions to tackle bias in AI translations. It was a brilliant way to collaborate with people from different fields, all combining their ideas.

Dr Antje Bothin (she, her), Hackathon participant



All 12 teams did a fantastic job of designing innovative solutions to bias in artificial intelligence. On the final day of the Hackathon, the jury chose five teams to present their ideas in more detail.



The Hackathon brought to front many interesting solutions sprung from the intensive hack sessions. Teams came up with algorithms to measure the ‘distance’ of bias in words susceptible to misgendering or bias; one team presented an idea of an easy gender toggle for post-translation editing. Some focused on the importance of creating more inclusive datasets for AI to train on, while others developed a sort of ‘bias-library’ for this purpose. Another team also developed a tool to allow the user to not only identify gender but also racial susceptibility of words to bias.

From these pitches, two winning teams were chosen to continue generating their solutions beyond the Hackathon. 
 

Looking forward

 
The issue of bias in AI is bigger than our teams could solve over a weekend, but, we have to start somewhere.
The fantastic range of solutions generated in such a short period of time shows just how possible it is to make the changes you want to see, given the right tools and time. In making AI more inclusive, all our teams helped bring us one step closer to helping AI embody everyone’s voice.

As for the winners of the Hackathon, they will continue to go on developing their solutions further, thanks to their €2500 cash prize, and support through mentoring and the resources of Goethe Institut.
 
We have still far to go, but each step counts: solutions from the hackathon, continued discussions and collaborative efforts, will help affect change in the way we and AI use our words; words that speak for everyone. 


Check out the 5 teams and their solutions to bias in AI shortlisted in our Artificially Correct Hackathon:

 
  • 1.	A “gender neutral” toggle. Team 2B generated a solution that focused on editing after machine translations. They introduced the idea of a “gender neutral” toggle which allows the user to choose, and decide what pronouns work best for them.   Given that translators work on a range of material, from literary texts to news features, this element of choice means you can pair machine translation outputs with the needs of the user.   Furthermore, giving the user agency to choose gender-fair and gender inclusive language empowers the user, and helps make non-binary people visible. © Goethe-Institut
    1. A “gender neutral” toggle
    Team 2B generated a solution that focused on editing after machine translations. They introduced the idea of a “gender neutral” toggle which allows the user to choose, and decide what pronouns work best for them. Given that translators work on a range of material, from literary texts to news features, this element of choice means you can pair machine translation outputs with the needs of the user. Furthermore, giving the user agency to choose gender-fair and gender inclusive language empowers the user, and helps make non-binary people visible.
  • 2.	Measuring gender bias. Next up, the Wheatstone Warriors prototyped a solution to measure gender bias. Using a data training set from Sweden, the team looked at how many masculine, feminine and gender-neutral pronouns there were, and tagged each one.   Tagging these genders generated a measurable ‘distance’ of bias which could be observed – the larger the gap, the more the phrase was susceptible to bias. They then began to fine tune the model, based on the phrases susceptibility to bias.     In achieving this, the team demonstrated that gender bias in data is not an insurmountable hurdle, but one that can be mitigated through tweaks after machines have been trained. © Goethe-Institut
    2. Measuring gender bias
    Next up, the Wheatstone Warriors prototyped a solution to measure gender bias. Using a data training set from Sweden, the team looked at how many masculine, feminine and gender-neutral pronouns there were, and tagged each one. Tagging these genders generated a measurable ‘distance’ of bias which could be observed – the larger the gap, the more the phrase was susceptible to bias. They then began to fine tune the model, based on the phrases susceptibility to bias. In achieving this, the team demonstrated that gender bias in data is not an insurmountable hurdle, but one that can be mitigated through tweaks after machines have been trained.
  • 3.	Crowdsourcing diverse data. Next onto the floor was team 1A, who set out to challenge data sets that lacked diversity. To promote diversity, the team created a platform that encouraged underrepresented groups to contribute speech data. The more diverse data, the more the computer would learn to recognise languages from all over the world, in all dialects – working towards making AI available to everyone, as well as representative.   The prototype draws users in using a mascot – in this case, your friendly neighbourhood egg. © Goethe-Institut
    3. Crowdsourcing diverse data
    Next onto the floor was team 1A, who set out to challenge data sets that lacked diversity. To promote diversity, the team created a platform that encouraged underrepresented groups to contribute speech data. The more diverse data, the more the computer would learn to recognise languages from all over the world, in all dialects – working towards making AI available to everyone, as well as representative. The prototype draws users in using a mascot – in this case, your friendly neighbourhood egg.
  • 4.	Unbiasing the world, together. BiasByUs created a website that acted as a ‘library’ for bias in translations. This collection of bias is to be generated through crowdsourcing bias found in machine translations. Collating these examples of bias from a variety of languages will ultimately help correct future AI training. Here, users are directly involved in helping counter bias, showing that together, we can unbias the world.   In addition to countering bias through data collection, the website also acts as a platform to raise awareness of bias in AI. © Goethe-Institut
    4. Unbiasing the world, together
    BiasByUs created a website that acted as a ‘library’ for bias in translations. This collection of bias is to be generated through crowdsourcing bias found in machine translations. Collating these examples of bias from a variety of languages will ultimately help correct future AI training. Here, users are directly involved in helping counter bias, showing that together, we can unbias the world. In addition to countering bias through data collection, the website also acts as a platform to raise awareness of bias in AI.
  • 5.	Identifying bias susceptibility. Team 4C focused on one of Google Translates, and other deep learning systems flaws - the struggle to comprehend complex content. It is in this complex content and context that words are more likely to be mistranslated.   As a solution the team developed a translation bias reduction tool that allows you to identify sentences susceptible to racial and gender bias. It works to identify and analyse the greater context of the sentence, and consequently highlights sentences and phrases that could be susceptible to bias.   This could also be used as a post-translation editing tool, highlighting where translators should focus their attention in the machine translated text to check for bias. © Goethe-Institut
    5. Identifying bias susceptibility
    Team 4C focused on one of Google Translates, and other deep learning systems flaws - the struggle to comprehend complex content. It is in this complex content and context that words are more likely to be mistranslated. As a solution the team developed a translation bias reduction tool that allows you to identify sentences susceptible to racial and gender bias. It works to identify and analyse the greater context of the sentence, and consequently highlights sentences and phrases that could be susceptible to bias. This could also be used as a post-translation editing tool, highlighting where translators should focus their attention in the machine translated text to check for bias.

Check out the video interviews with the winner teams discussing their projects and learnings from the hackathon: