$25,000 for those who succeed in choosing the right pronouns. Who’s in?
I recently had the opportunity to take part in a workshop on unconscious bias in artificial intelligence processing of natural language. This workshop was a first of its kind and was held as part of the annual conference of the Association of Computational Linguistics. As a researcher working on unconscious bias, it was an excellent opportunity for my team and I, BiaslyAI, to present what we’ve been working on for the past year. Other attendees included a group of enthusiastic researchers from industry and academia in the field of natural language processing, all working on mitigating gender bias in algorithms.proves a problematic task when multiple people are referenced in a text.
What is natural language processing?
Natural language processing is an interdisciplinary field that includes a mix of several disciplines such as artificial intelligence, linguistics, computer science etc. All textual data is studied under the aegis of natural language processing, which includes, for example, blog posts, news articles, question-and-answer formats, text-to-speech and even text generation. As algorithms have developed, a major problem has arisen: our unconscious biases show up in the results.
A $25,000 competition organized by Google
A competition was launched during the workshop: to develop algorithms so precise, they would be able to choose the right pronoun to use in a sentence. Google offered $25,000 to the winners.
Identifying the right pronoun is very important, and a particularly problematic task when one or more people are referenced in a text. Imagine a text where several individuals are mentioned. How is the algorithm supposed to know which pronoun to use? We also have to deal with the fact that some clichés or proper nouns tend to be associated with a “he” or a “she”. In this case, these generalizations become sexist prejudices.
Automatic translation from Turkish into English
Melvin Johnson, a workshop speaker and software developer at Google AI, is working on gender in machine translation using artificial intelligence. His discovery concerns the translation of Turkish into English.
The Turkish language is sometimes gender-neutral – neither feminine nor masculine – but when translated, it becomes gendered. Here’s an example: when the word “doctor” is ungendered in Turkish. When translated, it becomes “he is a doctor”. This is explained by the source of the data with which we have fed the algorithm. As explained earlier, our unconscious biases can influence the results of a translation.
The workshop was a great source of inspiration. With the rapid evolution of technology, it’s reassuring to know that many people around the world are striving to eliminate gender bias in computational linguistics.
