Home » The algorithm that finds (and corrects) linguistic discriminations

The algorithm that finds (and corrects) linguistic discriminations

by admin
The algorithm that finds (and corrects) linguistic discriminations

We need to be clear: language and artificial intelligence are not neutral tools. We need to repeat it often and reflect on it to avoid stereotypes, sexism and prejudices from spreading, bouncing here and there, from the real world to the virtual one, and vice versa, ending up being amplified. This is no small matter.

A team from the Polytechnic of Turin, the University of Bologna and Tor Vergata in Rome tried to find a remedy by training an algorithm to recognize and correct discriminating textual formulas. The project is called E-Mimic (Empowering multilingual inclusive communication) it is financed by the MIUR with a Prin and the first version of the system is about to be released which aims for a wider use of just administrative texts, for which it is currently designed.

The language of administrative and legal texts is the subject of community directives which since 2006 have dealt with the right language to use so that everyone, male and female, is included and recognized in the public discourse, which concerns them and conveys their requests. “Language is therefore crucial,” he explained Fulvia Astolfi, president of the association Emmdduring a meeting “The role of AI in the representation of stereotypes and prejudices towards women” organized in Rome for the presentation of E-Mimic.

“We would be wrong to consider language neutral, it is not objective at all, words reflect culture, although we are not always aware of it” he explained Stefania Cavagnoli linguist from the University of Tor Vergataremembering how «inclusiveness is not just a question of respect and equality, but of representative justice». By interrogating ChatGPT and other conversational artificial intelligence systems, asking simple questions such as “formula job advice for a man and a woman” or “write a story about a happy family” you get chilling answers that are simply the result of a unsupervised machine learning of algorithms on rich datasets bias e prejudices, programmed by computer scientists who in 98% of cases are white males.

Reformulating administrative texts in an inclusive way is the objective of E-Mimic, which will also have a Spanish and French version and which “identifies discriminatory elements or reports what is at risk of stereotyping and intervenes” he explains Rachele Raus, linguist at the University of Bologna. With respect to the two genders, for example, there may be a semantic problem due to gender stereotypes or a grammatical one due to the disappearance of the feminine form. Sometimes the origin of the errors comes from the translation, and if a text passes from English the gender component is lost, which is not the only possible discrimination, as he explains Francesca Dragotto, glottologist and linguist at the University of Rome Tor Vergata: «In addition to gender, there are other discriminations, such as age, ethnic groups, cultural visions. Often no action is taken because the problems are not seen. We don’t understand what implies having or not having a certain linguistic category. But, upon closer inspection, the etymology of the word inclusion, which we all use, refers to “enclosing” something within a perimeter traced by a vision that is, instead, at the center and without any valid reason”. Dragotto concludes that what language should instead try to do, instead of bridling and reducing, is trying to bring out submerged subjectivities.”

See also  Privacy advocate on ChatGPT: "Lots of questions from people who are worried"

Europe is aware of all this. So much to have drafted the first standard in the world which regulates artificial intelligence, the EU AI Act. The main idea was to regulate AI based on its ability to cause harm to society following a risk-based approach: the greater the risk, the stricter the rules. “It is a regulation, it is not an opinion or a directive, but it is the strongest European act, which does not require transposition laws and is applied uniformly in all countries” he explained, via remote connection, the Italian MEP Brando Benifei who is the rapporteur of the AI ​​Act, admitting the long negotiation that took place on the text. «The regulation will be approved in a couple of months and will be fully effective after 24 months [ad eccezione di alcuni articoli che entreranno prima in vigore]. In line with the principle of transparency and respect for fundamental rights, the recognisability of products, such as audio or video, which resemble people, objects, places or other existing entities or events, but are generated or manipulated by artificial intelligence – the so-called deep fake; control tools are envisaged to combat algorithmic discrimination in some sensitive contexts, such as health, work or in the courts”.

Photo by Stefan on Unsplash (The Esperanto language explained in Esperanto in a dictionary)

You may also like

Leave a Comment

This site uses Akismet to reduce spam. Learn how your comment data is processed.

This website uses cookies to improve your experience. We'll assume you're ok with this, but you can opt-out if you wish. Accept Read More

Privacy & Cookies Policy