AI researchers improve method to eliminate gender bias in natural language processing

According to a recent study this could be a key step in solving the problem of human bias creeping into artificial intelligence.

Although a computer itself is an unbiased machine, much of the data and programming that passes through computers is generated by humans. This can be a problem when conscious or unconscious human bias ends up being reflected in the text samples that AI models use to analyze and “understand” language.

Computers are not immediately able to understand text, explains Lei Ding, first author of the study and a graduate student in the Department of Mathematical and Statistical Sciences. They need the words to be converted into a set of numbers to understand them – a process called word integration.

“Natural language processing is basically teaching computers to understand texts and languages,” says Bei Jiangassociate professor in the Department of Mathematical and Statistical Sciences.

Once researchers have taken this step, they can then plot words as numbers on a 2D graph and visualize the relationships of the words to each other. This allows them to better understand the extent of gender bias and later determine whether the bias has actually been eliminated.

All the meaning, no bias

Although other attempts to reduce or remove gender bias from texts have been somewhat successful, the problem with these approaches is that gender bias is not the only thing removed from texts.

“In many gender bias reduction methods, when they reduce bias in a vector of words, they also reduce or eliminate important information about the word,” says Jiang. This type of information is known as semantic information and offers important contextual data that might be needed in future tasks involving these word incorporations.

For example, when considering a word like “nurse,” researchers want the system to remove any gender information associated with that term while retaining information that links it to related words such as doctor, hospital, and medicine.

“We need to preserve this semantic information,” Ding says. “Without this, the embeddings would perform very poorly. [in natural language processing tasks and systems].”

Fast, accurate — and fair

The new methodology also outperformed leading debiasing methods in various tasks assessed on word incorporation.

As it is refined, the methodology could offer a flexible framework that other researchers could apply to their own word incorporations. As long as a researcher has guidance on the correct group of words to use, the methodology can be used to reduce bias related to a particular group.

Although at this point the methodology still requires input from researchers, Ding explains that it might be possible in the future to have some kind of built-in system or filter that could automatically remove gender bias in a variety of contexts.

The new methodology is part of a larger project, titled BIAS: Accountable AI for Gender Equality and the Ethnic Labor Market, which seeks to solve real-world problems.

For example, people reading the same job posting may react differently to particular words in the description that often have a gender association. A system using the methodology created by Ding and his collaborators would be able to flag words that might change a potential candidate’s perception of the job or their decision to apply due to perceived gender bias, and suggest alternative words. to reduce this bias.

Although many AI models and systems focus on finding ways to perform tasks more quickly and accurately, Ding notes that teamwork is part of a growing field that seeks to make progress on another important aspect of these models and systems.

“People are focusing more on accountability and fairness in AI systems.”

The research was supported by the Canada-UK Artificial Intelligence Initiative.

James G. Williams