Using Interpretability Techniques and Deep Learning to Understand Nuances in Family Engagement

• In Event: AERA Poster Session 7
  In Poster Session: Technology, Instruction, Cognition and Learning SIG Poster Session 2 (San Diego)

Abstract

Machine learning techniques could be transformative for teacher education, but they also have the potential to reinforce societal biases and to provide inaccurate feedback to pre-service teachers (PSTs). Pre-trained classifiers, such as BERT, tend to be the most accurate for education datasets, but the models were not designed for education and can miss nuances or reinforce biases. Interpretable methods can support educators with fine-tuning pretrained machine learning models, provide new insights into data, and allow for bias monitoring. In this study, we tested the use of intpretable methods with the BERT model on transcripts from practice family conferences run by PSTs. We gained new insights into the capabilities of the model and how we can improve it for teacher education.

Authors

• Debajyoti Datta, University of Virginia

• Rose Sebastian, University of Virginia

• Laura Barnes