Snow White, Cinderella, and Sleeping Beauty have more in common than their origins as classic fairytale characters, and are now some of Disney’s most celebrated characters. According to literary scholars, their fairy tales are riddled with gender biases and stereotypes, and now include AI.
A team of researchers from Northeastern University, University of California, Los Angeles, and IBM Research developed an artificial intelligence framework that can analyze children’s picture books and detect instances of gender bias.
The way in which fairy tales portray and teach precepts, morals, and sociocultural roles to children, especially young girls, has been debated in academia and elsewhere for decades. These stories are filled with princesses in need of salvation and handsome princes who are there to save them.
Dacuo Wang, an associate professor at Northeastern University and one of the researchers, said tools like AI-driven spellchecking created by his team are being used by authors and publishers as well as researchers. , says it is expected to create more inclusive stories for children. project researchers.
“If I ever have a baby girl in the future, I hope she doesn’t get discouraged from taking on those jobs and overcoming those challenges.” [or] They say, ‘Someone will come help me,’ or, ‘You can’t do that as a girl,'” Wang said. “If we could develop a technology that would automatically detect and alert us to these kinds of gender biases and stereotypes, it would be a guardrail, at least not only for ancient fairy tales, but for the new stories that are being written and created every day today. It can act as a safety net.”
All of these studies began as part of the team’s ongoing research into how AI can help build language learning skills in young children. The team was already interested in fairy tales as a tool for language learning, collecting hundreds of stories from around the world to use as a “corpus” for their analytical algorithms.
They recruited a group of educational professionals, teachers and academics, to comb through the stories and develop a list of questions and answers that would help prove whether children are learning from these stories. bottom. Ultimately, 10,000 question-answer pairs were created, and all of these stories, wherever they came from, contained within them gender stereotypes. I noticed.
The princess will eat poisoned apples, be imprisoned, kidnapped, cursed, or die, and there is no way to change the situation. Meanwhile, the male characters – princes, kings and heroes – were slaying the dragon, breaking her curse and saving her princess.
Previous research in this area has focused on what Wang calls “surface-level” biases. That meant analyzing stories and identifying word and phrase combinations like “prince” and “brave” that connected ideas and identities in specific ways. But Wang and the rest of the team wanted to dig deeper.
They focused on the “chronological narrative event chain”, the specific combination and sequence of events and actions experienced or performed by characters.
“In fact, it is the experiences and actions that define who this person is, and what those actions affect the reader about. [they] Should I imitate that fictional character or shouldn’t I,” Wang said.
Using the hundreds of stories they collected, the team created an automated process to extract character names and genders along with every event. Then I arranged those events as a chain for each character. We also automated the process of grouping events and actions into specific categories. Each event was analyzed and given an odds ratio of how often it was associated with a male or female character.
Of the 33,577 events analyzed in the study, 69% were due to male characters and 31% were due to female characters. Events associated with female characters were often associated with domestic chores such as grooming, cleaning, cooking, and sewing, while those associated with male characters were associated with failure, success, or aggression.
With all this information, Wang and team created a natural language processing tool that can not only analyze individual events, but also spot biases within event chains.
“Some were saved, got married and lived happily ever after. Others killed monsters, saved princesses and lived happily ever after,” Wang said. “It’s not the ‘lived happily’ part or the ‘getting married’ part that’s different. In fact, it’s the chain of events that preceded these events that makes the difference.”
By automating this process, Wang said he hopes the tool will be leveraged among those outside the research community who are actually creating or recreating these stories. I’m here. In the process, we can begin to stop stories from passing these outdated and harmful ideas to the next generation.
“With our tool, just upload your first draft into a tool like this and it should generate a score or meter that says, ‘There are things you want to check and things you don’t want to check here. is. If this intent is not what you want to express, you may need to consider rewriting it. Here are some suggestions,” Wang said.
Going forward, Wang and team plan to expand the study to investigate other forms of bias. They also plan to use the tool to assess other AI biases. They want to use algorithms to analyze whether ChatGPT has same-gender biases and stereotypes when it creates content based on these stories.
“We are proposing that this is really a challenge, a challenge that the tech community can actually overcome,” Wang said. “I’m not saying our method is the best. I’m just saying that our method is the first to do this task, and this task is very dominant. …Maybe , maybe we should shift some of our attention to these existing social issues and challenges.”
Cody Mello-Klein is a reporter for Northeastern Global News.send him an email c.mello-klein@northeastern.edu. follow him on twitter @Proelectioneer.
