Leveraging Natural Language Processing to Guide Educators
Natural language processing (NLP) is a component of AI that allows a computer program to understand human language as it’s spoken or written. In his M-Powering Teachers project, which is funded by a Grand Challenges Team Project Grant from the University of Maryland (UMD), Assistant Professor of Education Policy Jing Liu and his team are using NLP to analyze transcripts from K-12 mathematics classrooms and provide specific, actionable feedback to help teachers improve. For example, they’re exploring how often the educators interact with students or take up their ideas.
“I generate the feedback as a way [to guide] teachers’ professional learning, so that, for schools and classrooms that don't have a lot of resources, and teachers who don't have a lot of support in terms of their instructional practices, they can at least rely on AI tools,” says Liu, who is a member of the Institute for Trustworthy AI in Law & Society (TRAILS.) “I think that's just another way to think about equity.”
One of his team’s initial observations is that NLP alone appears to be less helpful to educators than NLP combined with human input. Liu noted that teachers who were given AI-based analyses alone generally found it challenging to interpret the feedback and weren’t particularly motivated to do so. However, when combined with human coaching, NLP seemed to be a very helpful training tool.
In fact, Liu recently received a grant from the Overdeck Family Foundation to conduct a study that will assess the impact of combining human coaching with automated feedback.
He is also leading an effort jointly supported with $4.5 million from the Bill & Melinda Gates Foundation, the Walton Family Foundation and the Chan Zuckerberg Initiative to create high-quality benchmark data on math teaching from diverse upper elementary and middle schools—information that is currently lacking, he said.
“From a technical perspective, when you develop AI models, the initial data you use to train the models has to be pretty representative, so that your downstream model and applications can be more unbiased,” he added.
To address that, Liu is collecting a wealth of information from demographic surveys and other sources about students’ test scores, sense of belonging, perceptions of mathematics and more. The data will be used to train subsequent AI models using the highest-quality—and most equitable—information possible. “So rather than just … developing more AI models, let’s take a step back and create the best data first,” he explained.
—This is an excerpt of a larger story on AI in the UMD College of Education written by Christina Folz for Endeavors magazine
