U.S. math students still aren’t making the grade on the world’s stage, but a team of researchers backed by the National Science Foundation is putting AI to the test for improving the teaching of the subject in public schools.
Teachers in two Colorado school districts in the spring started pilot tests with AI for analyzing math class discussions. Participating teachers helped with the design of the program as well as testing.
The goal is to improve engagement in math classes by analyzing teachers’ use of discussion techniques and students’ responses, and providing the instructors feedback for follow-up.
The program is already showing promising signs, with teachers saying it offers useful insights.
“I’m always trying to improve conversations that happen in my math lessons and to help students have discussions with each other to explain their thinking. The application lets me see how I’m doing at meeting that goal,” said Kristin Holmquist, a fifth grade teacher in the pilot tests at Eagleview Elementary School.
Accelerating STEM Education
The U.S. remains an underperformer globally for math education. Its math students rank 31st in the world, according to the Organization for Economic Cooperation and Development.
Legislators, universities, companies and influencers alike are supporting STEM — science, technology, engineering and math — education.
University of Colorado Boulder researchers developed the classroom application, dubbed Talk Moves, using natural language processing models run on NVIDIA GPUs. Talk Moves taps speech recognition to automatically generate classroom transcripts and natural language processing models to analyze the text for discussion insights. The app provides teachers feedback on their use of specific discourse techniques known in math education circles as “talk moves.”
“There’s a big push to have teachers think deeply about their discourse practices and the conversations they have in math classes with students,” said Jennifer Jacobs, an associate research professor at the university’s Institute of Cognitive Science.
Teaching NLP on GPUs
The researchers gathered more than 500 written transcripts of K-12 math lessons to train the Talk Moves system. The transcripts were annotated for six different types of talk moves used in sentences, totalling more than 200,000 sentences for the training dataset. Annotating the dataset was handled by two talk moves language specialists for two years.
The university researchers fine-tuned the Bidirectional Encoder Representations from Transformers (BERT) natural language processing model on the Talk Moves dataset. They preprocessed the data on local NVIDIA GPUs and then trained the model on cloud instances of NVIDIA GPUs.
The parallel processing capabilities and Tensor Core architecture of NVIDIA GPUs enable higher throughput and scalability for working with complex language models — offering record-setting performance for both the training and inference of BERT.
Training their hefty BERT-based model on NVIDIA GPUs enabled quicker iterations of the app, said Abhijit Suresh, a graduate research assistant in the institute.
“We use the GPU parallelization to make sure we can train the model much faster — it’s much faster than running it on CPUs,” said Suresh.
The resulting Talk Moves classifier model is used to predict the label for which discussion technique, or talk move, is applied in class.
AI for Student Equity
The effort comes amid a bipartisan legislative proposal in the U.S. Senate to modernize math teaching in the nation’s public schools.
Research on how to integrate AI in the classroom to best support teachers and students is now being extended as part of a $20 million research collaboration led by the University of Colorado and backed by the NSF, intended to boost STEM learning opportunities for students from historically underrepresented populations.
Talk Moves aims to broaden and deepen classroom math conversations, supporting student equity, said Jacobs. The Talk Moves app has a series of NVIDIA GPU-driven classifiers that can measure how often students, and which students, are responding in discussions.
“A big goal of accountable talk is equity, because we want all students listening, participating, talking and being a part of that community,” she said.
The research team — Abhijit Suresh, Jennifer Jacobs, Vivian Lai, Chenhao Tan, Karla Scornavacco, Wayne Ward, James Martin and Tamara Sumner — recently submitted a paper on their work.
NVIDIA NGC enables remote training of BERT models.